Paper Atlas
Paper Library
Explore papers on object pose estimation and related areas. Structured metadata connects methods and research directions with paper details, code reproduction, and experimental records.
Each paper records its research area, publication details, datasets, evaluation metrics, author affiliations, code availability, and method tags.
Research areas distinguish broad topics such as visual localization, feature matching, open-vocabulary perception, VLA, and world models. Method tags provide finer filtering by technical approach.
Filters
No filters selected
235 papers
3D-Object Perception Transformer (3PT)
3PT unifies detection, segmentation, and 6-DoF pose with two Transformers for 2D / 3D object-centric understanding. Continuous refinement of object representations and rendered-feature alignment support zero-shot cross-object perception on BOP-H3, BOP-Industrial, and BOP-Classic.
Agastya Kalra, Tim Salzmann, Guy Stoppi, Dmitrii Marin, Rishav Agarwal, Vage Taamazyan, Martin Bokeloh, Stefan Hinterstoisser, Anton Boykov, Alberto Dall’Olio, Pravin Dangol, Kartik Venkataraman, Huaijin Chen
Area: Object Pose Estimation
Datasets: BOP-H3 / BOP-Industrial / BOP-Classic
Metrics: 2D detection AP / segmentation AP / 6DoF pose AP-mm / BOP ranking
Affiliations: Intrinsic
6DAttack: Backdoor Attacks in the 6DoF Pose Estimation
Studies backdoor attacks on PnP-based and end-to-end 6-DoF pose pipelines. Structurally 3D triggers and modified pose labels are injected into training data so that models behave normally on clean inputs but produce controlled incorrect poses when triggered.
Jihui Guo, Zongmin Zhang, Zhen Sun, Yuhao Yang, Jinlin Wu, Fu Zhang, Xinlei He
Area: Object Pose Estimation
Datasets: LINEMOD / YCB-Video / CO3D
Metrics: ADD-C / ADD-P / PEA / 2DPE / ASR
Affiliations: The University of Hong Kong / HKUST (Guangzhou) / Beihang University / Chinese Academy of Sciences
Active Tactile Exploration for Rigid Body Pose and Shape Estimation
A violation-implicit loss incorporates physical-constraint violations into joint estimation of freely moving unknown convex rigid-body trajectories and cuboid / 20-vertex polytope shapes. Observed / Fisher information at a local MLE estimates EIG to select the next touch. In simulation, the EIG strategy yields stable errors below 4 mm after 5 cuboid actions; real Trifinger results at step 5 substantially outperform random selection. Convex geometry, local optima, multiple initializations, and tactile / proprioceptive errors limit the method.
Ethan K. Gordon, Bruke Baraki, Hien Bui, Michael Posa
Area: Object Pose Estimation
Datasets: Drake simulated cuboid and Platonic-solid trials / self-collected Trifinger DenseTact trials
Metrics: bidirectional Chamfer distance in mm / actions since first contact / 95% confidence interval / two-tailed t-test
Affiliations: University of Pennsylvania GRASP Laboratory
Aleatoric Uncertainty from AI-based 6D Object Pose Predictors for Object-relative State Estimation
The paper adds independently calibratable heteroscedastic MLPs to PoET's translation and rotation branches. Gaussian NLL predicts six log-variance components as dynamic measurement covariance for a MaRS EKF. Dynamic covariance alone performs similarly to fixed noise; consistent benefits come from uncertainty-based anchor switching and outlier rejection. Mean position RMSE over 7 real flights decreases from 0.172 m to 0.146 m. Symmetric-object rotation coverage is poorly calibrated, and translation–rotation correlations and multimodality are omitted.
Thomas Jantos, Stephan Weiss, Jan Steinbrener
Area: Object Pose Estimation
Datasets: synthetic power-pole inspection dataset / synthetic YCB-Video subset / AIVIO real UAV inspection dataset
Metrics: prediction interval coverage probability at 95% / AUC of ADD-S / translation and rotation error / position and orientation RMSE / maximum position error / TensorRT inference latency
Affiliations: University of Klagenfurt, Control of Networked Systems Group
AlignPose: Generalizable 6D Pose Estimation via Multi-view Feature-metric Alignment
AlignPose converts candidate poses from multiple calibrated RGB images into a shared world frame, then jointly optimizes one SE(3) pose in frozen visual feature space. Discrepancies between rendered-object and observed features drive generalizable refinement.
Anna Šárová Mikeštíková, Médéric Fourmy, Martin Cifka, Josef Sivic, Vladimir Petrik
Area: Object Pose Estimation
Datasets: YCB-V / T-LESS / HouseCat6D / ITODD-MV / IPD / XYZ-IBD
Metrics: BOP AR / VSD / MSSD / MSPD
Affiliations: Czech Institute of Informatics, Robotics and Cybernetics, Czech Technical University in Prague
Learning 6D Object Pose Estimation With Event Cameras Using Synthetic Data and Domain Randomization
AEE transforms RGB templates from 4,000 hemispherical viewpoints × 10 in-plane angles into an event latent codebook through edge extraction and domain randomization. Real event crops retrieve rotations, while box centers / object scale approximate translation. Retrieval takes around 5 ms, but full-pose 10% ADD-S is highly sensitive to detection boxes: only 2.5% for the wood block versus 86.0% for the cracker box.
Oussama Abdul Hay, Xiaoqian Huang, Muhammad Ahmed Humais, Abdulla Ayyad, Randa Almadhoun, Yahya Zweiri
Area: Object Pose Estimation
Datasets: E-POSE
Metrics: ADD-S at 10% diameter for rotation / ADD-S at 10% diameter for full pose / inference latency
Affiliations: Advanced Research and Innovation Center / University of Sunderland / Khalifa University
Breaking the 3D Dataset Bottleneck: Fast Scalable Generation of Aligned 3D Assets from Scratch for Category 6D Pose Estimation and Robotic Grasping
This work mass-produces canonically aligned object assets and complete annotations through a text-to-image-to-3D pipeline. It supports category-level 6D pose, shape completion, and robotic grasping while reducing the time and manual effort required for high-quality 3D data.
Duret Guillaume, Danylo Mazurak, Florence Zara, Jan Peters, Liming Chen
Area: Object Pose Estimation
Datasets: GenOmni3D / category-level pose data / robot grasping data
Metrics: pose consistency / shape completion IoU / grasp success / generation time
Affiliations: Inria / TU Darmstadt
C3PO: Canonicalization of 3D Pose from Partial Views with Generalizable Correspondence Features
C3PO formulates partial-view category canonicalization as cross-video correspondence and registration. Features generalizable to unseen categories support pairwise registration and global optimization to align multiple partial shapes.
Yu Chi, Leonhard Sommer, Olaf Dünkel, Dominik Muhle, Daniel Cremers, Christian Theobalt, Adam Kortylewski
Area: Object Pose Estimation
Datasets: ShapeNet / CO3Dv2
Metrics: 30° canonicalization accuracy
Affiliations: Technical University of Munich / University of Freiburg / Max Planck Institute for Informatics / CISPA Helmholtz Center for Information Security
CanoVerse: 3D Object Scalable Canonicalization and Dataset for Generation and Pose
CanoVerse replaces manual continuous SO(3) canonicalization of 3D objects with 5 candidates generated from category templates and geometric / semantic criteria, followed by one-click human selection. From 750,000 Objaverse / Objaverse-XL assets, it retains 320,000 canonicalized assets across 1,156 categories. VI-Net's cross-domain OmniObject3D orientation recovery reaches Acc@10° 20.18 and Acc@30° 24.83 versus 6.56 / 9.60 with Objaverse-OA training. It estimates rotation of centered complete point clouds relative to canonical frames, without camera-frame translation or scene-level RGB/RGB-D detection, segmentation, and full 6D pose.
Li Jin, Yuchen Yang, Weikai Chen, Yujie Wang, Dehao Hao, Tanghui Jia, Yingda Yin, Zeyu Hu, Runze Zhang, Keyang Luo, Li Yuan, Long Quan, Xin Wang, Xueying Qin
Area: Object Pose Estimation
Datasets: CanoVerse / Objaverse / Objaverse-XL / OmniObject3D / Toys4K / Google Scanned Objects / DREDS
Metrics: Acc@10 degrees / Acc@30 degrees / absolute angular error / Chamfer Distance / LPIPS / CLIP score / orientation IQR / Recall@10 / Recall@30 / IC / GEC / annotation seconds per object
Affiliations: Shandong University / LIGHTSPEED STUDIOS / University of North Carolina at Chapel Hill / Hong Kong University of Science and Technology / Peking University
CAOA: Completion-Assisted Object-CAD Alignment
CAOA first completes missing object geometry in scans using context-aware point completion. Symmetry encoding and a Siamese object–CAD alignment network then predict 9-DoF alignment: translation, rotation, and per-axis scale.
Hiranya Garbha Kumar, Minhas Kamal, Balakrishnan Prabhakaran
Area: Object Pose Estimation
Datasets: Scan2CAD / ScanNetv2 / S2C-Completion / SN-Indoor
Metrics: 9DoF alignment accuracy / Chamfer Distance
Affiliations: University at Albany, State University of New York
COG: Confidence-aware Optimal Geometric Correspondence for Unsupervised Single-reference Novel Object Pose Estimation
COG formulates single-reference unknown-object pose estimation as optimal-transport correspondence with pointwise confidence. Geometry, semantics, and cycle consistency suppress nonoverlapping regions and outliers; weighted SVD then recovers relative 6-DoF pose.
Yuchen Che, Jingtu Wu, Hao Zheng, Asako Kanezaki
Area: Object Pose Estimation
Datasets: TUD-L / LM-O / YCB-V / BOP
Metrics: BOP AR / VSD / MSSD / MSPD / overlap IoU / inference time
Affiliations: Institute of Science Tokyo / RIKEN / Tohoku University
Color-Pair Guided Robust Zero-Shot 6D Pose Estimation and Tracking of Cluttered Objects on Edge Devices
Relatively illumination-stable edge color pairs support both initial RGB-D–mesh registration and optical-flow correspondence filtering. Mean ADD / ADD-S across 15 textured YCB-V objects is 60.9 / 78.6, with Fast-YCB tracking ADD of 93.13. Five-object initialization runs at around 7 FPS on Jetson Orin, and tracking exceeds 20 FPS on Xavier, but texture, CAD, detection, and depth are required.
Xingjian Yang, Ashis G. Banerjee
Area: Object Pose Estimation
Datasets: YCB-Video / Fast-YCB / LoCoBot real-world sequence
Metrics: ADD / ADD-S / tracking ADD / relative rotation error / relative translation error / optical-flow EPE / FPS
Affiliations: University of Washington Department of Mechanical Engineering / University of Washington Department of Industrial & Systems Engineering
ComPose: A Unified Completion-Pose Framework for Robust Category-Level Object Pose Estimation
ComPose unifies shape completion and pose estimation for category-level partial point clouds. It predicts sparse keypoints and local dense neighborhoods, then uses geometric relation encoding and a consistency loss for NOCS alignment, reducing dependence on category shape priors.
Huan Ren, Yihan Chen, Chuxin Wang, Nailong Liu, Wenfei Yang, Tianzhu Zhang
Area: Object Pose Estimation
Datasets: REAL275 / CAMERA25
Metrics: 10 degree 2 cm accuracy / 3D IoU / inference FPS
Affiliations: University of Science and Technology of China / Beijing Institute of Control Engineering
ConceptPose: Training-Free Zero-Shot Object Pose Estimation using Concept Vectors
ConceptPose uses an LLM to generate open-vocabulary concepts, then a VLM produces dense saliency and 3D concept maps. Concept-vector similarity establishes cross-view 3D–3D correspondences, and Umeyama alignment solves zero-shot relative pose.
Liming Kuang, Yordanka Velikova, Mahdi Saleh, Jan-Nico Zaech, Danda Pani Paudel, Benjamin Busam
Area: Object Pose Estimation
Datasets: REAL275 / Toyota-Light / YCB-Video / LINEMOD
Metrics: ADD(-S) recall / ADD-AUC / ADD-S-AUC / BOP AR / 3D IoU
Affiliations: University of Bonn / Toyota Research Institute
CoordAR: One-Reference 6D Pose Estimation of Novel Objects via Autoregressive Coordinate Map Generation
For single-reference 6D pose estimation of unknown objects without CAD, reference / query 3D–3D correspondences are represented as an ROC map. Autoregressive generation of discrete tokens replaces continuous coordinate regression, and Umeyama alignment recovers the relative pose.
Dexin Zuo, Ang Li, Wei Wang, Wenxian Yu, Danping Zou
Area: Object Pose Estimation
Datasets: Real275 / Toyota-Light / YCB-Video / LINEMOD / FoundationPose / OO3D-9D
Metrics: ADD(-S) Recall / AUC ADD / AUC ADD-S / AR (MSSD/MSPD/VSD) / Inference time
Affiliations: Shanghai Jiao Tong University / Midea Group
Cov2Pose: Leveraging Spatial Covariance for Direct Manifold-aware 6-DoF Object Pose Estimation
Cov2Pose compresses image features into SPD matrices through spatial covariance pooling. A manifold-aware pose head encodes rotation and translation, addressing inconsistencies of direct Euclidean regression under occlusion and across rotation representations.
Nassim Ali Ousalah, Peyman Rostami, Vincent Gaudillière, Emmanuel Koumandakis, Anis Kacem, Enjie Ghorbel, Djamila Aouada
Area: Object Pose Estimation
Datasets: LINEMOD / Occluded LINEMOD / YCB-Video
Metrics: ADD(-S) / AUC ADD-S / AUC ADD(-S) / inference time
Affiliations: University of Luxembourg
DICArt: Advancing Category-level Articulated Object Pose Estimation in Discrete State-Spaces
DICArt places category-level articulated-object pose in a discrete state space. A Flexible Flow Decider improves discrete diffusion denoising, while hierarchical kinematic coupling jointly recovers rigid parts, joint axes, and articulation states under occlusion.
Li Zhang, Mingyu Mei, Ailing Wang, Xianhui Meng, Yan Zhong, Xinyuan Song, Liu Liu, Rujing Wang, Zaixing He, Cewu Lu
Area: Object Pose Estimation
Datasets: ReArtMix / RobotArm / synthetic articulated objects
Metrics: rotation error / translation error / joint angle error / axis distance error
Affiliations: University of Science and Technology of China / Shanghai Jiao Tong University / Zhejiang University / East China Normal University / Peking University / Emory University / Hefei University of Technology
DKPMV: Dense Keypoints Fusion from Multi-View RGB Frames for 6D Pose Estimation of Textureless Objects
DKPMV samples 512 CAD keypoints with FPS, improves per-view predictions through symmetry-aware training and local attention, then reconstructs keypoint clouds from dense multiview correspondences, aligns with Umeyama, and optimizes reprojection. With 8 RealSense views on ROBI, overall ADD AR / 5 mm 10° recall is 95.5 / 93.6. CAD, calibrated extrinsics, and target detection remain required.
Jiahong Chen, Jinghao Wang, Zi Wang, Ziwen Wang, Banglei Guan, Qifeng Yu
Area: Object Pose Estimation
Datasets: ROBI
Metrics: ADD average recall / 5 mm 10° recall / 2 mm 3° recall
Affiliations: National University of Defense Technology / Hunan Provincial Key Laboratory of Image Measurement and Vision Navigation
DynOPETs: A Versatile Benchmark for Dynamic Object Pose Estimation and Tracking in Moving Camera Scenarios
DynOPETs provides 175 RGB-D sequences with moving objects / cameras, around 210,000 frames, and over 530,000 annotations, covering category-level and unseen-object pose estimation and tracking. Annotations fuse FoundationPose absolute poses, TAP relative motion, and an SE(3) pose graph. PGO mean ATE is 4.94 mm on 10 additional motion-capture sequences. Most ground truth still comes from a pseudo-label pipeline with manually removed graph edges; independent motion-capture validation covers only 10 sequences.
Xiangting Meng, Jiaqi Yang, Mingshu Chen, Chenxin Yan, Yujiao Shi, Wenchao Ding, Laurent Kneip
Area: Object Pose Estimation
Datasets: DynOPETs COPE-119 / DynOPETs UOPE-56 / ten-object marker-based mocap validation set
Metrics: 3D IoU / 5°2 cm and 10°5 cm accuracy / BOP average recall / ATE / rotation and translation RPE / AUC of ADD and ADD-S
Affiliations: ShanghaiTech University / Fudan University
EgoXtreme: A Dataset for Robust Object Pose Estimation in Egocentric Views under Extreme Conditions
EgoXtreme evaluates egocentric object pose under low light, motion blur, smoke, severe occlusion, and close-range truncation. Industrial maintenance, sports, and emergency-rescue scenarios test the real-world generalization of laboratory methods.
Taegyoon Yoon, Yegyu Han, Seojin Ji, Jaewoo Park, Sojeong Kim, Taein Kwon, Hyung-Sin Kim
Area: Object Pose Estimation
Datasets: EgoXtreme / YCB-V / HOT3D
Metrics: ADD(-S) / ADD-S AUC / pose tracking accuracy / condition-wise recall
Affiliations: Seoul National University / VGG, University of Oxford
Keypoint-Based Dynamic Object 6-DoF Pose Tracking via Event Camera
The network detects semantic keypoints from event time surfaces, builds 2D–3D correspondences through a precomputed hash, and automatically initializes with EPnP. Polarity, spatial distance, and local event density propagate keypoints, while an EKF suppresses jitter. Most rotation errors on custom simulated / EVK4 data are substantially below EDOPT's, but data are limited and both models and keypoints are known beforehand.
Zhe Wang, Qijin Song, Zihao Li, Jingyu Xiao, Weibang Bai
Area: Object Pose Estimation
Datasets: self-collected simulated mechanical-parts event set / self-collected Prophesee EVK4 real event set
Metrics: relative rotation error in deg/s / relative translation error in cm/s
Affiliations: ShanghaiTech Automation and Robotics Center / ShanghaiTech University / WLSA Shanghai Academy
Event-based Motion & Appearance Fusion for 6D Object Pose Tracking
This learning-free event tracker decodes event optical flow and speed-independent EROS in parallel. Flow propagates 6D velocity through a Kalman filter; EROS corrects pose by matching 13 local renderings, followed by UKF smoothing. Mean ablation errors across 8 synthetic sequences are 2.18 cm / 14.57°, approaching FoundationPose on some fast motions. Real data lack ground truth, the complete online system is not implemented, and around 110 Hz is inferred only from component timings.
Zhichao Li, Chiara Bartolozzi, Lorenzo Natale, Arren Glover
Area: Object Pose Estimation
Datasets: synthetic 500 Hz YCB motion sequences / self-collected real YCB event sequences
Metrics: translation RMSE and standard deviation in cm / rotation RMSE and standard deviation in degrees / component latency
Affiliations: Istituto Italiano di Tecnologia / University of Genoa
Event6D: Event-based Novel Object 6D Pose Tracking
Event6D targets high-speed tracking of unknown objects. Event streams and the latest depth frame reconstruct intensity / depth at arbitrary timestamps for render-and-compare tracking. The system exceeds 120 FPS and evaluates generalization on real event data.
Jae-Young Kang, Hoonhee Cho, Taeyeop Lee, Minjun Kang, Bowen Wen, Youngho Kim, Kuk-Jin Yoon
Area: Object Pose Estimation
Datasets: EventBlender6D / EventHO3D / Event6D
Metrics: pose tracking accuracy / ADD(-S) / tracking FPS / rotation / translation error
Affiliations: KAIST / NVIDIA
Fast-WAM: Do World Action Models Need Test-time Future Imagination?
Fast-WAM retains future-video prediction during training but uses attention masks so actions read only current observations. At deployment, Wan2.2 acts as a single-pass encoder followed by 10-step action-expert denoising, without generating future video. Mean success is 91.8% / 97.6% on RoboTwin 2.0 / LIBERO. Latency on RTX 5090D V2 is 190 ms versus IDM's 810 ms, around 4.26× faster. This world-action model requires robot demonstrations and does not output object 6D pose; real tasks still trade speed against success.
Tianyuan Yuan, Zibin Dong, Yicheng Liu, Hang Zhao
Area: World Models
Datasets: LIBERO / RoboTwin 2.0 / Galaxea R1 Lite towel folding
Metrics: task success rate / average completion time / inference latency
Affiliations: IIIS, Tsinghua University / Galaxea AI
Finding NeMO: A Geometry-Aware Representation of Template Views for Few-Shot Perception
NeMO encodes a few RGB templates into object-centric point memory with geometry and semantics. A shared decoder reuses this memory for detection, segmentation, surface prediction, and 6-DoF pose of objects unseen in training.
Sebastian Jung, Leonard Klüpfel, Rudolph Triebel, Maximilian Durner
Area: Object Pose Estimation
Datasets: BOP HOPEv2 / HANDAL / T-LESS / TUD-L / YCB-Video
Metrics: AP / AR / MSPD / MSSD
Affiliations: German Aerospace Center (DLR)
Follow My Hold: Hand-Object Interaction Reconstruction through Geometric Guidance
FollowMyHold adds multistage geometric guidance to sampling from a pretrained rectified-flow 3D generator. Joint optimization of hand / object transforms, constrained by depth, normals, silhouettes, keypoints, and SDF contact, produces physically consistent hand–object interaction reconstructions.
Ayce Idil Aytekin, Helge Rhodin, Rishabh Dabral, Christian Theobalt
Area: Object Pose Estimation
Datasets: OakInk / ARCTIC / DexYCB
Metrics: F-score@5mm / F-score@10mm / Chamfer Distance / intersection volume / reconstruction rate
Affiliations: Max Planck Institute for Informatics / Saarland University / Bielefeld University
GaussianArt: Unified Modeling of Geometry and Motion for Articulated Objects
GaussianArt jointly learns part geometry and motion decomposition with articulated 3D Gaussians. Part segmentation, L0 regularization, and trajectory regularization support digital twins of articulated objects with up to 20 parts.
Licheng Shen, Saining Zhang, Honghan Li, Peilin Yang, Zihao Huang, Zongzheng Zhang, Hao Zhao
Area: Object Pose Estimation
Datasets: MPArt-90 / PartNet-Mobility / MultiScan
Metrics: axis angle error / axis position error / part motion error / Chamfer Distance
Affiliations: Beijing Academy of Artificial Intelligence / Tsinghua University / Nanyang Technological University / Beijing Institute of Technology / Huazhong University of Science and Technology
GenHOI: Generalized Hand-Object Pose Estimation with Occlusion Awareness
GenHOI jointly recovers MANO hands and object rotation / translation from object-centered monocular RGB. InstructBLIP generates object, hand, and interaction descriptions; multimodal masked reconstruction over images, text, and training ground-truth mesh point clouds learns occlusion-aware representations. Predicted hand rotation and 3D joints constrain object rotation and translation heads. On three unseen DexYCB S3 objects, reported mean ADD-0.5D is 89.34 and ADD-S is 21.76 mm, well above the listed comparisons. All four hand metrics slightly trail UniHOPE, and HO3Dv2 PA-MJE is not best. Inference needs neither CAD nor a depth sensor, but requires object crops and image-guided voxel sampling; 18 FPS excludes offline InstructBLIP text generation.
Hui Yang, Wei Sun, Jian Liu, Jian Xiao, Tao Xie, Hossein Rahmani, Ajmal Mian, Nicu Sebe, Gim Hee Lee
Area: Object Pose Estimation
Datasets: DexYCB S0 / DexYCB S3 / HO3Dv2
Metrics: ADD-0.5D / ADD-S / MJE / PA-MJE / ST-MJE / V-PE / OCE / MCE / OME / F@5 / F@15
Affiliations: NERC-RVC, Hunan University / National University of Singapore / Lancaster University / The University of Western Australia / University of Trento
GeniWorld: A Generalizable Interactive World Model for Robotic Manipulation via Visual Actions
GeniWorld renders URDF-based numerical actions as camera-aligned visual motion conditions, driving a Wan2.2 causal video model to predict environmental responses for closed-loop policy evaluation and data synthesis. On RoboTwin Clean-to-Random it reports FVD 20.15 and EWMScore 63.54. Synthetic data raise aggregate policy success across four real tasks and five settings from 40.8% to 69.0%. It generates future observations rather than actions or object poses, requires robot models and camera calibration, and the official code is still marked Coming soon.
Chenghao Gu, Hanyang Yu, Jingbo Zhang, Haitao Lin, Wenyao Zhang, Jinghe Wang, Hanglei Jin, Shuzhao Xie, Jingyan Jiang, Zhi Wang
Area: World Models
Datasets: RoboTwin 2.0 / WorldArena evaluation protocol / Xtrainer real-world manipulation
Metrics: LPIPS / PSNR / SSIM / FID / FVD / EWMScore / task success rate
Affiliations: Shenzhen International Graduate School, Tsinghua University / Tencent Robotics X / The Hong Kong University of Science and Technology / Shenzhen Technology University
GraspClutter6D: A Large-scale Real-world Dataset for Robust Perception and Grasping in Cluttered Scenes
GraspClutter6D contains 1000 dense real scenes, 200 objects, 52,000 RGB-D images, 735,545 pose instances, and 9.3 billion analytical grasp candidates. A rigid four-camera UR5 platform, 13-view fusion, and review by two experts establish poses. Object-level force-closure grasps are transformed into scenes and collision-filtered. Training data raise real grasp success in 15-object piles from GraspNet-1B's 54.9% to 67.9%. The 9.3 billion grasps are candidates, not physical executions, and depth residuals combine pose, mesh, and sensor errors.
Seunghyeok Back, Joosoon Lee, Kangmin Kim, Heeseon Rho, Geonhyup Lee, Raeyoung Kang, Sangbeom Lee, Sangjun Noh, Youngjin Lee, Taeyeop Lee, Kyoobin Lee
Area: Object Pose Estimation
Datasets: GraspClutter6D / YCB-Video transfer benchmark / real UR5 cluttered-grasp benchmark
Metrics: mean absolute depth discrepancy / viewpoint coverage / real grasp success rate / ADD(-S) and ADD-S AUC / instance segmentation AP / 6-DoF grasp AP
Affiliations: Korea Institute of Machinery and Materials / Gwangju Institute of Science and Technology / Korea Advanced Institute of Science and Technology
Reconstructing Hand-Held Objects in 3D from Images and Videos
MCC-HO jointly recovers hand–object geometry from a single image and an inferred 3D hand. RAR retrieves a matching 3D object model with a VLM and rigidly aligns it to multiframe observations. The hand provides a strong geometric anchor for scale and position.
Jane Wu, Georgios Pavlakos, Georgia Gkioxari, Jitendra Malik
Area: Object Pose Estimation
Datasets: DexYCB / MOW / HOI4D / 100 Days of Hands
Metrics: F-score@5mm / F-score@10mm / Chamfer Distance / intersection volume
Affiliations: University of California, Berkeley / The University of Texas at Austin / California Institute of Technology
HIPPo: Harnessing Image-to-3D Priors for Model-Free Zero-Shot 6D Pose Estimation
HIPPo builds a complete metric mesh from the first RGB-D frame using Wonder3D + MASt3R in around 8 seconds, then estimates / tracks with FoundationPose at around 15 FPS. A viewpoint sphere triggers replacement of diffusion priors with real observations. Single-reference YCB-V / LM-O ADD AUC is 89.07 / 88.49. Rapid onboarding is the benefit, while mesh updates take around 9.87 seconds and heavy occlusion can corrupt initial generation.
Yibo Liu, Zhaodong Jiang, Binbin Xu, Guile Wu, Yuan Ren, Tongtong Cao, Bingbing Liu, Rui Heng Yang, Amir Rasouli, Jinjun Shan
Area: Object Pose Estimation
Datasets: YCB-Video / LM-O / HIPPo custom reconstruction set
Metrics: ADD AUC / ADD-S AUC / Chamfer distance / initialization and update latency
Affiliations: Huawei Noah’s Ark Lab / York University / University of Toronto
Towards in-the-wild Egocentric 3D Hand-Object Pose Estimation
EPIC-Contact expands 2,272 egocentric stable-grasp clips into 62.3K frames of posed hand–object meshes through hand-contact painting, region-level bijective contact transfer, and EC-fit. HOPformer's 12-layer decoder lets DINOv2 object tokens read WiLoR hand-pose tokens, jointly predicting bimanual MANO, object classes, rotation / translation, and optional joint angles in one pass. Against the strongest JointTransformer, EPIC CDev drops from 30.1 to 20.7 mm and SR@0.05 rises from 17.6% to 29.8%. Main pose tables use oracle meshes, automatic classification accuracy is only 52.9%, and the official repository reports approximately 1.28× projection-scale and 0.78× absolute-depth biases in EPIC checkpoints. Results mainly establish relative / root-aligned hand–object improvements, rather than open-world or calibrated absolute 6D pose.
Siddhant Bansal, Zhifan Zhu, Shashank Tripathi, Jiahe Zhao, Michael J. Black, Dima Damen
Area: Object Pose Estimation
Datasets: EPIC-Contact / ARCTIC / EPIC-Grasps / EPIC-KITCHENS VISOR
Metrics: CDev / MRRPE / MDev / hand / object ACC / MPJPE / AAE / SR@0.05 / SR@0.1 / object classification accuracy
Affiliations: University of Bristol / Max Planck Institute for Intelligent Systems, Tübingen
HRDexDB: A Paired Human-Robot Dataset for Cross-Embodiment Dexterous Grasping
HRDexDB records 100 objects and 2.1K paired human / four-dexterous-hand grasping sequences with 21 external and 2 egocentric cameras. Around 24M frames include 3D hands, object 6D poses, robot states, and tactile annotations. Contact transfer, cross-embodiment grasp retrieval, hand reconstruction, and object pose under grasp occlusion benchmark its uses.
Jongbin Lim, Taeyun Ha, Mingi Choi, Jisoo Kim, Byungjun Kim, Subin Jeon, Hanbyul Joo
Area: Robot Perception & Manipulation
Datasets: HRDexDB / FreiHAND / OmniRobotHome / Google Scanned Objects
Metrics: grasp success rate / R@1 / R@3 / R@5 / PA-MPJPE / PA-MPVPE / ADD / ADD-S / AR_MSSD
Affiliations: Seoul National University / RLWRLD
IndustryShapes: An RGB-D Benchmark Dataset for 6D Object Pose Estimation of Industrial Assembly Components and Tools
IndustryShapes builds Classic and Extended BOP datasets from 5 textureless, reflective, thin-structure, or symmetric industrial parts: 34 scenes, around 22,400 pose annotations, and 10 static RGB-D onboarding sequences. FoundationPose model-based AR on Classic / Extended is 0.67 / 0.69 versus model-free 0.25 / 0.33. Limits include only 5 objects, uneven train / test scene complexity, and inconsistent Extended depth-error reporting between the text and Table IV.
Panagiotis Sapoutzoglou, Orestis Vaggelis, Athina Zacharia, Evangelos Sartinas, Maria Pateraki
Area: Object Pose Estimation
Datasets: IndustryShapes Classic / IndustryShapes Extended
Metrics: ADD recall / BOP average recall / VSD / MSSD / MSPD / detection and segmentation mAP / depth discrepancy
Affiliations: National Technical University of Athens
InstantPose: Zero-Shot Instance-Level 6D Pose Estimation From a Single View
InstantPose generates a coarse mesh and 400 templates from one unposed RGB reference. DINOv2 matches the query, depth from both sides lifts 3D correspondences, and RANSAC + Umeyama jointly recover rotation, translation, and scale. YCB-V BOP AR is 36.83 at around 3.9 s on an A4500. Single-view onboarding is convenient, but reconstruction quality, occlusion, and weak-texture matching limit results.
Francesco Di Felice, Alberto Remus, Stefano Gasperini, Benjamin Busam, Lionel Ott, Stefan Thalhammer, Federico Tombari, Carlo Alberto Avizzano
Area: Object Pose Estimation
Datasets: YCB-Video
Metrics: BOP average recall / VSD / MSSD / MSPD / inference time / robot grasp success
Affiliations: Scuola Superiore Sant’Anna / Leonardo S.p.A. / Technical University of Munich / MCML / ETH Zürich Autonomous Systems Lab / UAS Technikum Wien / Google Zürich
InvariantCloud: A Globally Invariant, Uniquely Indexed Point Cloud Framework for Robust 6-DoF Tactile Pose Tracking
InvariantCloud prebuilds a global tactile mesh of 475 uniquely identified points, allowing direct per-frame matching without LK or nearest-neighbor search. Kabsch solves roll / pitch, contact-contour PCA solves yaw, and centroid / height recover translation. One-minute static translation MAE is 0.22 / 0.14 / 0.11 mm, with yaw 0.84°. It outperforms NormalFlow on weak contours such as eggs, but yaw is unobservable for perfectly spherical contact.
Pengfei Ye, Yuxiang Ma, Yi Zhou, Wei Chen, Wenzhen Dong, Molong Duan
Area: Object Pose Estimation
Datasets: self-collected GelSight Mini household-object sequences / self-collected multi-contact tactile SLAM sequence
Metrics: one-minute static cumulative MAE / return-pose repeatability error / Z-axis tracking error / surface reconstruction consistency
Affiliations: Hong Kong University of Science and Technology / Massachusetts Institute of Technology / Chinese University of Hong Kong
iTACO: Interactable Digital Twins of Articulated Objects from Casually Captured RGBD Videos
From RGB-D video containing handheld camera motion, object motion, and occlusion, iTACO obtains coarse joint states from static / dynamic correspondences, then optimizes joint parameters and articulated parts through differentiable refinement and part segmentation.
Weikun Peng, Jun Lv, Cewu Lu, Manolis Savva
Area: Object Pose Estimation
Datasets: iTACO S-Dataset / iTACO L-Dataset / PartNet-Mobility
Metrics: joint axis error / joint position error / joint type error / joint state error / Chamfer Distance
Affiliations: Simon Fraser University / Shanghai Jiao Tong University
KV-Tracker: Real-Time Pose Tracking with Transformers
KV-Tracker caches global information from a multiview visual geometry Transformer as a key–value scene representation, supporting online object-pose and geometry updates in monocular video while reducing repeated global computation and long-term drift.
Marwan Taher, Ignacio Alzugaray, Kirill Mazur, Xin Kong, Andrew Davison
Area: Object Pose Estimation
Datasets: TUM RGB-D / object pose tracking benchmarks
Metrics: tracking accuracy / ATE / rotation / translation error / FPS
Affiliations: Imperial College London
M-VTOP: Modular Visuo-Tactile Object Pose Estimation for High-Precision Robotic Manipulation
M‑VTOP converts vision and GelSight observations into masks, rendering CAD masks for each SE(3) particle and updating belief by IoU. Object–environment signed distance adds a soft-contact likelihood. Across five real insertion categories, fused vision / touch achieves mean success of 93% / 100% with SAM2 / Mask R-CNN, above individual modalities. The full system runs at only 0.4 Hz and assumes CAD, calibration, rigid objects, and stable grasps.
Miquel Oller, Qiyang Qian, Radu Corcodel, Siddarth Jain
Area: Object Pose Estimation
Datasets: five-object simulated pose benchmark / five-object real insertion benchmark
Metrics: mean bounding-box pose error / translation error in mm / rotation error in degrees / insertion success rate
Affiliations: Mitsubishi Electric Research Laboratories / University of Michigan / University of California, Berkeley
MemPose: Category-level Object Pose Estimation with Memory
MemPose adds per-category external feature memory of length 96 to an AG-Pose-style RGB-D category-level 9-DoF pipeline. Historical pooled local keypoint features are updated through similarity merging with Gumbel perturbations; current keypoints read memory through attention and gated fusion. Main results freeze memory: REAL275 IoU75 81.0 / 5°2 cm 59.9, HouseCat6D IoU50 81.5 / IoU75 56.4, and Wild6D IoU75 46.2. The model uses 225M parameters at 33 FPS versus AG-Pose's 223M at 35 FPS. Test-time updates raise REAL275 5°2 cm to 63.6 only in additional experiments, without a complete stream-order or contamination-prevention protocol.
Xiao Lin, Minghao Zhu, Yun Peng, Liuyi Wang, Qiyi Wang, Chengju Liu, Qijun Chen
Area: Object Pose Estimation
Datasets: REAL275 / CAMERA25 / HouseCat6D / Wild6D
Metrics: 3D IoU mAP / 5° 2cm / 5° 5cm / 10° 2cm / 10° 5cm
Affiliations: Tongji University / State Key Laboratory of Autonomous Intelligent Unmanned Systems
Monocular Vehicle Pose and Shape Reconstruction via Dynamic Context Adaptation and Progressive Geometry Refinement
For distant vehicles in autonomous driving, MonoVPR's HDCA fuses object-centric geometry with scene-centric semantics. BIMR then uses tanh-bounded iterative mesh refinement to improve both 3D pose and shape recovery.
Wei Li, Long Ji, Ying Wang, Xiao Wu, Zhaoquan Yuan, Penglin Dai
Area: Object Pose Estimation
Datasets: ApolloCar3D
Metrics: A3DP-Abs / A3DP-Rel / Rotation Error / Translation Error / mIoU
Affiliations: Southwest Jiaotong University / Engineering Research Center of Sustainable Urban Intelligent Transportation
Deep Learning-Based Object Pose Estimation: A Comprehensive Survey
A systematic review of instance-level, category-level, and novel-object pose estimation. Its evaluation section separates 3-DoF, 6-DoF, 9-DoF, and shape reconstruction, covering geodesic errors, BOP, ADD(-S), 2D projection, 3D IoU, and Chamfer distance.
Jian Liu, Wei Sun, Hui Yang, Zhiweng Zeng, Chongpei Liu, Jin Zheng, Xingyu Liu, Hossein Rahmani, Nicu Sebe, Ajmal Mian
Area: Object Pose Estimation
Datasets: BOP / LINEMOD / YCB-Video / NOCS / HouseCat6D
Metrics: Geodesic rotation error / ADD / ADD-S / ADD(S) / ADD AUC / n° mcm / 2D Projection / VSD / MSSD / MSPD / BOP AR / 3D IoU / Chamfer Distance
Affiliations: National Engineering Research Center for Robot Visual Perception / Lancaster University / University of Western Australia
OrienPose: Orientation-Guided Novel View Synthesis for Single-Image Unseen Object Pose Estimation
For single-image novel-object view synthesis, OrienPose introduces Orientation-Aware Guidance and an Orientation Consistency Loss to constrain geometry across view changes. Matching against synthesized-view templates then estimates pose.
Yating Liu, Zhaoshuai Qi, Yang Zou, Yongnan Yang, Shizhou Zhang, Yanning Zhang
Area: Object Pose Estimation
Datasets: ShapeNet / NAVI
Metrics: pose accuracy / ADD(-S) / AUC / ACC@30
Affiliations: Northwestern Polytechnical University
Pose Anything Anywhere: Model-free Object Poses from Arbitrary References
PANY estimates CAD-free unseen-object 6D pose using a canonically posed anchor and optional unposed assist views. LoRA-adapted VGGT predicts multiview geometry; fused geometric–visual descriptors, cross-view contrastive learning, and pose-graph registration recover query pose. RGB uses PnP, while RGB-D uses Umeyama. Single-reference YCB-V ADD AUC is 92.5, rising to 94.6 with 8 assist views; LINEMOD ADD-0.1d rises from 55.3 to 87.4. Arbitrary / pose-free references still assume an anchor with a defined canonical pose, object masks, and sufficient view connectivity.
Hongli Xu, Jiaqi Hu, Junwen Huang, Boyang Zhong, Peter KT Yu, Nassir Navab, Benjamin Busam, Slobodan Ilic
Area: Object Pose Estimation
Datasets: Omni6DPose / REAL275 / Toyota-Light / YCB-Video / LINEMOD / LINEMOD-Occlusion
Metrics: BOP Average Recall / ADD-0.1d / ADD AUC / inference time
Affiliations: Technical University of Munich / Munich Center for Machine Learning / Siemens AG / XYZ Robotics / ROBOX
Point2Pose: Occlusion-Recovering 6D Pose Tracking and 3D Reconstruction for Multiple Unknown Objects Via 2D Point Trackers
Point2Pose starts from a few clicks per unknown rigid object and RGB-D video. SAM2 supplies masks, a long-range 2D point tracker maintains associations through occlusion, and depth-lifted tracks populate object-centric keypoint maps. Multi-hypothesis SVD / RANSAC, TSDF scoring, and SDF refinement estimate per-frame poses; factor graphs and online TSDF preserve global consistency and reconstruction. YCBMultiTrack-Real-World ADD-S / ADD AUC is 89.40 / 71.19 versus BundleSDF's successful-sequence means of 62.08 / 42.34. Ordinary averages on HO3D and YCBInEOAT do not uniformly lead. Texture, accurate masks, RGB-D, and a static camera are required; outputs are first-frame-relative rather than CAD-canonical poses.
Tzu-Yuan Lin, Ho Jae Lee, Kevin Doherty, Yonghyeon Lee, Sangbae Kim
Area: Object Pose Estimation
Datasets: HO3D v3 / YCBInEOAT / YCBMultiTrack-Synthetic / YCBMultiTrack-Real-World
Metrics: ADD-S AUC / ADD AUC / Chamfer Distance
Affiliations: MIT Biomimetic Robotics Lab / Massachusetts Institute of Technology / Boston Dynamics
Pose-VLA: Universal Pose Pretraining for Generalizable Vision-Language-Action Policies
Pose-VLA uses camera-centric discrete pose tokens to unify nonrobot 3D annotations with robot end-effector trajectories. Spatial grounding / motion alignment is learned from 1.4 million images, 6.5 million 3D annotations, and around 1.55 million trajectories, followed by a flow-matching action expert for few-shot embodiment alignment. Results reach 79.91% / 79.10% on RoboTwin 2.0 Easy / Hard, mean 96.0% on LIBERO, and mean 81.25% across four real bimanual tasks. Pose prediction is only a pretraining proxy; deployment does not explicitly output pose. Depth and raymaps may be masked for RGB-only inference.
Haitao Lin, Hanyang Yu, Jingshun Huang, He Zhang, Yonggen Ling, Ping Tan, Xiangyang Xue, Yanwei Fu
Area: VLA / Embodied AI
Datasets: Omni3D / Omni6DPose / BOP Challenge / AgibotWorld Beta / InternData-A1 / RoboTwin 2.0 / LIBERO
Metrics: 3D grounding AP15 / task success rate
Affiliations: Tencent Robotics X / Futian Laboratory / The Hong Kong University of Science and Technology / Fudan University / Shanghai Innovation Institute
PoseGAM: Robust Unseen Object Pose Estimation via Geometry-Aware Multi-View Reasoning
PoseGAM processes a query image and multiple posed templates with a geometry-aware multiview network. Explicit point geometry and learned geometry features directly predict pose without explicitly constructing query–template correspondences, improving average AR across five BOP datasets.
Jianqi Chen, Biao Zhang, Xiangjun Tang, Peter Wonka
Area: Object Pose Estimation
Datasets: LM-O / T-LESS / TUD-L / IC-BIN / YCB-V
Metrics: BOP AR / ARA / ATA / AUC
Affiliations: King Abdullah University of Science and Technology
PoseGaussian: 6D Pose Estimation for Unseen Objects via Sparse-View Object-Level 3D Gaussian Splatting
PoseGaussian reconstructs an object-level 3D Gaussian Splatting representation from sparse RGB-D reference views. Learned ICP initialization and differentiable rendering refinement with appearance and depth consistency estimate unknown-object 6D pose.
Wubin Shi, Shaoyan Gai, Feipeng Da
Area: Object Pose Estimation
Datasets: LINEMOD / GenMOP / real-world object scenes
Metrics: ADD(S)-0.1d / AR / rotation / translation error / inference time
Affiliations: Southeast University
Exploring Category-level Articulated Object Pose Tracking on SE(3) Manifolds
Tracks category-level articulated-object 6D poses across frames by learning pose increments in SE(3) / se(3). Weighted PPF voting, dynamic keyframes, and kinematic joint constraints reduce drift and improve consistency between parts.
Xianhui Meng, Yukang Huo, Li Zhang, Liu Liu, Haonan Jiang, Yan Zhong, Pingrui Zhang, Cewu Lu, Jun Liu
Area: Object Pose Estimation
Datasets: PM-Videos / ReArt-Videos / RobotArm-Videos / PartNet-Mobility / ReArt-48 / RobotArm
Metrics: Rotation error / Translation error / 3D IoU / Inference time
Affiliations: University of Science and Technology of China / China Agricultural University / Hefei University of Technology / Zhejiang University of Technology / Fudan University / Shanghai AI Laboratory / Peking University / Shanghai Jiao Tong University
SCAPO: Self-Supervised Category-Level Articulated Pose Estimation from a Single 3D Observation
SCAPO combines an SE(3)-equivariant autoencoder, joint-aware blend skinning, cycle reconstruction, and instance residuals to jointly recover canonical geometry, parts, and joint states from one 3D observation. Synthetic experiments require no GT labels; the real-data setting uses segmentation and joint supervision without pose supervision.
Can Zhang, Gim Hee Lee
Area: Object Pose Estimation
Datasets: HOI4D / Shape2Motion / real articulated-object RGB-D scans
Metrics: rotation error / translation error / joint angle error / part segmentation
Affiliations: National University of Singapore
SE(3)-PoseFlow: Estimating 6D Pose Distributions for Uncertainty-Aware Robotic Manipulation
SE(3)-PoseFlow encodes RGB-D with DINOv2 + PointNet++, conditions SE(3) rectified flow through DiT* masked cross-attention, and directly samples multimodal poses. DBSCAN, Chamfer distance, or neural SDFs select a pose. REAL275 10°5 cm is 89.1%, and YCB-V / LM-O is 68.2% / 53.7%. Multiple hypotheses raise occluded-cup grasp success from 6/10 to 9/10, with costs from 50 samples, mesh reranking, and interpreting samples in Bayesian terms.
Yufeng Jin, Niklas Funk, Vignesh Prasad, Zechu Li, Mathias Franzius, Jan Peters, Georgia Chalvatzaki
Area: Object Pose Estimation
Datasets: REAL275 / YCB-Video / LINEMOD-Occlusion / real Franka mug-grasp trials
Metrics: 5°2 cm accuracy / 5°5 cm accuracy / 10°2 cm accuracy / 10°5 cm accuracy / FPS / real-robot grasp success
Affiliations: TU Darmstadt / Honda Research Institute Europe / DFKI / Hessian.AI
SE(3)-Equivariance with Geometric and Topological Guidance for Category-Level Object Pose Estimation
SEGPose combines SE(3)-equivariant feature prediction with geometric and topological guidance, using local point cloud structures and global connectivity to strengthen category-level pose estimation. It also evaluates transfer to real robotic grasping.
Sheng Yu, Di-Hua Zhai, Yuanqing Xia
Area: Object Pose Estimation
Datasets: category-level pose benchmarks / robotic grasping scenes
Metrics: rotation error / translation error / 3D IoU / grasp success
Affiliations: Beijing Institute of Technology / Zhongyuan University of Technology
Category-Level Object Shape and Pose Estimation in Less Than a Millisecond
The paper formulates category-level shape–pose MAP conditions as a nonlinear eigenproblem over unit quaternions. SCF iterations require only a minimum eigenvector of a 4×4 matrix per step for local solutions, and Lagrangian duality checks global optimality. A synthetic solve takes around 0.12 ms and the REAL275 solver around 1.26 ms, excluding roughly 50 ms of frontend keypoint detection.
Lorenzo Shaikewitz, Tim Nguyen, Luca Carlone
Area: Object Pose Estimation
Datasets: synthetic active-shape benchmark / CAST / NOCS REAL275 / ApolloCar3D
Metrics: solver runtime / rotation error / translation error / A3DP
Affiliations: MIT Laboratory for Information and Decision Systems / Boston University / MIT SPARK Lab
Simultaneous Extrinsic Contact and In-Hand Pose Estimation via Distributed Tactile Sensing
TacGraph combines dual-GelSight tactile geometry, in-grasp displacement, and wrench estimates in a factor graph. Geometric consistency, nonpenetration, contact kinematics, and force balance jointly estimate in-grasp rest pose, external contact points, and forces. Tactile-only contact-point error drops from ICP's 17.53 mm to 3.72 mm, with 23/40 successful open-loop insertions at 3 mm clearance. Known object / environment meshes, elastic nonslipping grasps, and single-point contact are assumed.
Mark Van der Merwe, Kei Ota, Dmitry Berenson, Nima Fazeli, Devesh K. Jha
Area: Object Pose Estimation
Datasets: self-collected GelSight contact-pose benchmark / self-collected tactile-only peg-insertion benchmark
Metrics: ADD in mm / contact-point error in mm / contact-force error in N / peg-insertion success rate
Affiliations: University of Michigan / Mitsubishi Electric / Mitsubishi Electric Research Laboratories
TwinTrack: Bridging Vision and Contact Physics for Real-Time Tracking of Unknown Objects in Contact-Rich Scenes
TwinTrack asynchronously reconstructs Gaussian Splats / neural SDFs from RGB-D in Real2Sim and uses contact trajectories to jointly correct collision geometry and physical properties. Online Sim2Real adaptively fuses one-step physics predictions according to visual matching quality. Across 8 scenes, mean ADD / ADD-S drops from the visual baseline's 3.01 / 1.15 cm to 1.23 / 0.59 cm at over 20 Hz. Ground truth is FoundationPose pseudo-labels, and prolonged visual absence still causes drift.
Wen Yang, Zhixian Xie, Yiting Wang, Abhijit Tadepalli, Heni Ben Amor, Shan Lin, Wanxin Jin
Area: Object Pose Estimation
Datasets: self-collected impact-rich falling sequences / self-collected LEAP in-hand manipulation sequences
Metrics: ADD in cm / ADD-S in cm / one-step contact-dynamics prediction error / tracking frequency
Affiliations: Arizona State University / Mohamed bin Zayed University of Artificial Intelligence
VPHO: Joint Visual-Physical Cue Learning and Aggregation for Hand-Object Pose Estimation
For hand–object interaction from a single RGB image, VPHO learns hand / object heatmaps and physical cues such as contact forces. It generates multiple diffusion pose hypotheses, then combines visual and physics-based aggregation to choose a visually consistent, physically feasible result.
Jun Zhou, Chi Xu, Kaifeng Tang, Yuting Ge, Tingrui Guo, Li Cheng
Area: Object Pose Estimation
Datasets: DexYCB / HO3Dv2 / DexYCB Phy / HO3Dv2 Phy
Metrics: MJE / PA-MJE / MCE / OCE / ADD-S / CP / PD / SD
Affiliations: China University of Geosciences / University of Alberta
WALDO: Where Unseen Model-based 6D Pose Estimation Meets Occlusion
Within a SAM6D-style model-based novel-object pipeline, WALDO uses coarse correspondences for both multiple initial poses and pointwise occlusion / background probabilities. Probability-guided dynamic nonuniform dense sampling, repeated sparse-to-dense matching, and weighted SVD refine poses. On five BOP-Core* sets, UAR / AR is 0.56 / 0.73 versus SAM6D's 0.54 / 0.71, with image runtime dropping from 4.37 s to 1.53 s. ICBIN shows the largest UAR gain, 0.50→0.56, but absolute UAR remains below FRTPose, FreeZe, and Co-op. UAR averages across visibility deciles to expose conventional AR's dominance by highly visible samples. RGB-D, exact CAD, detection / segmentation, and a complex multistage system remain required.
Sajjad Pakdamansavoji, Yintao Ma, Amir Rasouli, Tongtong Cao
Area: Object Pose Estimation
Datasets: LM-O (LMO) / T-LESS / TUD-L / IC-BIN / YCB-Video / BOP-Core* five-dataset subset
Metrics: Unbiased Average Recall (UAR) / BOP Average Recall (AR) / VSD recall / MSSD recall / MSPD recall / visibility-decile mAPD / mARD / per-image inference runtime
Affiliations: Huawei Technologies Canada
XYZ-IBD: Benchmarking Robust 6D Object Pose Estimation under Real-World Industrial Complexity
XYZ-IBD mounts RealSense, DLP structured-light, and Photoneo sensors on a FANUC arm to capture 75 dense bin scenes of 15 reflective, textureless, highly symmetric industrial parts. The ECCV-accepted version reports 22.5K frames and around 273K pose instances. Depth captured with antireflection spray, multiview fusion, manual coarse registration, and multiscale ICP establish labels. BOP 6D detection AP is only 0.266 for GDRN, 0.247 for SurfEmb, 0.564 for FoundationPose, and 0.578 for SAM6D, well below BOP-Core 5. Reported 0.999 mm / 0.432° accuracy comes from simulated error quantification reproducing noise and annotation steps. The project page retains an older 91K-annotation count; use v3 / ECCV statistics and identify the version.
Junwen Huang, Jiaqi Hu, Peter KT Yu, Slobodan Ilic, Martin Sundermeyer, Benjamin Busam
Area: Object Pose Estimation
Datasets: XYZ-IBD / BOP Core: LM-O / BOP Core: T-LESS / BOP Core: YCB-V / BOP Core: IC-BIN / BOP Core: TUD-L
Metrics: BOP 2D detection AP / BOP 6D detection AP / MSSD / MSPD / annotation translation error / annotation angular error / viewpoint calibration RMSE
Affiliations: Technical University of Munich / Munich Center for Machine Learning / XYZ Robotics / ROBOX / Google
You Only Pose Once: A Minimalist's Detection Transformer for Monocular RGB Category-Level 9D Multi-Object Pose Estimation
YOPO outputs categories, 2D boxes, centers, depth, rotation, and anisotropic scale directly from DETR object queries. Translation is conditioned on boxes, and pose error enters Hungarian matching. A fine-tuned model reaches REAL275 IoU50 79.6 and 10°10 cm 54.1. Its benefit is end-to-end RGB-only estimation; fine-grained accuracy still trails strong RGB-D systems.
Hakjin Lee, Junghoon Seo, Jaehoon Sim
Area: Object Pose Estimation
Datasets: CAMERA25 / REAL275 / HouseCat6D
Metrics: 3D IoU at 50% / 75% / 5°2 cm / 5°5 cm / 10°2 cm / 10°5 cm / 10°10 cm / latency
Affiliations: PIT IN Co. AI Robot Team
6DOPE-GS: Online 6D Object Pose Estimation using Gaussian Splatting
Incremental 2D Gaussian Splatting jointly optimizes object pose and representation. Dynamic keyframe selection, graph optimization, and pruning based on opacity statistics support online model-free RGB-D 6D tracking and reconstruction.
Yufeng Jin, Vignesh Prasad, Snehal Jauhri, Mathias Franzius, Georgia Chalvatzaki
Area: Object Pose Estimation
Datasets: HO3D / YCBInEOAT
Metrics: ADD / ADD-S / rotation error / translation error / tracking speed
Affiliations: Technische Universität Darmstadt / Honda Research Institute Europe / Hessian.AI
Any6D: Model-free 6D Pose Estimation of Novel Objects
Any6D learns joint object alignment and scale from one RGB-D anchor, then uses render-and-compare to select and refine pose / size hypotheses for unknown objects.
Taeyeop Lee, Bowen Wen, Minjun Kang, Gyuree Kang, In So Kweon, Kuk-Jin Yoon
Area: Object Pose Estimation
Datasets: REAL275 / Toyota-Light / HO3D / YCBInEOAT / LM-O
Metrics: AUC ADD / AUC ADD-S / BOP AR / AR_VSD / MSSD / MSPD
Affiliations: KAIST / NVIDIA
Betsu-Betsu: Multi-View Separable 3D Reconstruction of Two Interacting Objects
Betsu-Betsu represents interaction scenes with a shared multiresolution hash grid, decoding separate object SDFs and radiance fields. Alpha-blending regularization suppresses geometric interpenetration at mutual-occlusion boundaries.
Suhas Gopal, Rishabh Dabral, Vladislav Golyanik, Christian Theobalt
Area: 3D Representations / Reconstruction
Datasets: Betsu-Betsu human-object dataset / WildRGB-D / AffordPose / NeuralDome / ReMoCap
Metrics: Chamfer Distance / F-score@1cm / PSNR / SSIM / LPIPS / 3D IoU
Affiliations: Saarland University / Max Planck Institute for Informatics
BoxDreamer: Dreaming Box Corners for Generalizable Object Pose Estimation
Reconstructs an object's 3D bounding box from sparse reference RGB images, predicts its 2D box corners in the query view, and estimates unseen-object 6D pose with PnP, emphasizing generalization under sparse views and occlusion.
Yuanhong Yu, Xingyi He, Chen Zhao, Junhao Yu, Jiaqi Yang, Ruizhen Hu, Yujun Shen, Xing Zhu, Xiaowei Zhou, Sida Peng
Area: Object Pose Estimation
Datasets: LINEMOD / Occlusion LINEMOD / YCB-Video / OnePose / OnePose-LowTexture / Objaverse
Metrics: ADD / ADD-S / ADD(s)-0.1d / Proj2D@5px / ADD-AUC / ADD-S AUC
Affiliations: Zhejiang University / Xiangjiang Laboratory / Ant Group / EPFL / Chongqing University / Northwestern Polytechnical University / Shenzhen University
CAP-Net: A Unified Network for 6D Pose and Size Estimation of Categorical Articulated Parts from a Single RGB-D Image
CAP-Net jointly predicts part classes, center offsets, and NPCS with a single-stage RGB-D network, recovering pose, size, and joint states of category-level articulated parts on RGBD-Art.
Jingshun Huang, Haitao Lin, Tianyu Wang, Yanwei Fu, Xiangyang Xue, Yi Zhu
Area: Object Pose Estimation
Datasets: RGBD-Art / synthetic articulated-object benchmarks
Metrics: part AP50 / 5° 5cm / rotation error / translation error / size error
Affiliations: Fudan University / Huawei Noah’s Ark Lab
CapeLLM: Support-Free Category-Agnostic Pose Estimation with Multimodal Large Language Models
Query images and textual keypoint descriptions replace conventional support images and annotations. MLLM visual-language tokens, dynamic round training, and flexible coordinate decoding support category-agnostic keypoint pose estimation; this is not BOP-style 6D pose estimation.
Junho Kim, Hyungjin Chung, Byung-Hoon Kim
Area: Object Pose Estimation
Datasets: MP-100
Metrics: PCK@0.2 / PCK@0.1 / 1-shot accuracy / 5-shot accuracy
Affiliations: EverEx / Yonsei University
A Unified End-to-End Network for Category-Level and Instance-Level Object Pose Estimation from RGB Images
CIPE unifies category-level and instance-level RGB pose as set prediction, introduces a learnable Euclidean rotation representation, and injects category point cloud priors into bounding-box queries through prior–query fusion. The public GitHub repository is currently a placeholder without runnable code.
Jiale Ren, Hong Liu, Jinfu Liu, Peifeng Jiang
Area: Object Pose Estimation
Datasets: Category-level RGB pose benchmark / Instance-level RGB pose benchmark
Metrics: rotation / translation pose accuracy / category-level and instance-level protocol
Affiliations: Peking University
CleanPose: Category-Level Object Pose Estimation via Causal Learning and Knowledge Distillation
Front-door adjustment mitigates dataset bias in category-level pose estimation, while residual knowledge distillation from a 3D foundation model supplies less biased semantics. RGB-D 9-DoF pose and size estimation evaluates cross-instance generalization.
Xiao Lin, Yun Peng, Liuyi Wang, Xianyou Zhong, Minghao Zhu, Yi Feng, Jingwei Yang, Chengju Liu, Qijun Chen
Area: Object Pose Estimation
Datasets: REAL275 / CAMERA25 / HouseCat6D
Metrics: 3D IoU / 5° 2cm / 5° 5cm / 10° 2cm / 10° 5cm
Affiliations: Tongji University / State Key Laboratory of Autonomous Intelligent Unmanned Systems
Co-op: Correspondence-based Novel Object Pose Estimation
Co-op establishes semi-dense correspondences using a small number of CAD-rendered templates. A hybrid classification-plus-offset representation, probabilistic flow, and differentiable PnP enable coarse-to-fine 6D pose estimation of unknown objects.
Sungphill Moon, Hyeontae Son, Dongcheol Hur, Sangwook Kim
Area: Object Pose Estimation
Datasets: BOP: LM-O / T-LESS / TUD-L / IC-BIN / ITODD / YCB-V / HB
Metrics: BOP AR / AR_VSD / AR_MSSD / AR_MSPD / runtime
Affiliations: NAVER LABS
CRISP: Object Pose and Shape Estimation with Test-Time Adaptation
CRISP jointly estimates implicit shape and pose-normalized points, turning an active shape model corrector into a certifiable pseudo-label generator. Test-time self-training addresses large domain gaps.
Jingnan Shi, Rajat Talak, Harry Zhang, David Jin, Luca Carlone
Area: Object Pose Estimation
Datasets: YCB-Video / SPE3R / NOCS-REAL275
Metrics: ADD-S / ADD-S AUC / Chamfer distance / mAP
Affiliations: MIT Laboratory for Information & Decision Systems (LIDS)
Deterministic Object Pose Confidence Region Estimation
Deterministic Gaussian keypoint regression and inductive conformal prediction first construct 2D confidence regions. The implicit function theorem propagates them into 6D pose regions, avoiding the latency and region inflation of sampling-based uncertainty quantification.
Jinghao Wang, Zhang Li, Zi Wang, Banglei Guan, Yang Shang, Qifeng Yu
Area: Object Pose Estimation
Datasets: Occlusion LINEMOD / SPEED
Metrics: coverage rate / confidence region volume / ADD(-S) / runtime
Affiliations: National University of Defense Technology / Hunan Provincial Key Laboratory of Image Measurement and Vision Navigation
Distilling 3D Distinctive Local Descriptors for 6D Pose Estimation
dGeDi distills FreeZe's pointwise, local-reference-frame-dependent GeDi descriptors into PTV3. Ground-truth poses transfer CAD query teacher features to noisy partial targets, and a piecewise loss prioritizing low errors suppresses less salient teacher points. Geometry-feature extraction drops from 3.284 s to 0.019 s while average pose AR on five BOP datasets is maintained or slightly improved.
Amir Hamza, Andrea Caraffa, Davide Boscaini, Fabio Poiesi
Area: Feature Matching / Correspondences
Datasets: MegaPose-ShapeNetCore / LM-O / T-LESS / TUD-L / IC-BIN / YCB-V
Metrics: Ratio of Nearest points / Feature Matching Recall / BOP Average Recall / feature extraction runtime
Affiliations: Fondazione Bruno Kessler / University of Trento
High-Precision Object Pose Estimation Using Visual-Tactile Information for Dynamic Interactions in Robotic Grasping
A particle filter fuses visual and tactile displacements: vision tracks the contacted object in real time, while touch corrects slippage and pose changes of the held object. The abstract evaluates dynamic adaptation on 13 objects of different geometries but gives no cm / degree errors.
Zicai Peng, Te Cui, Guangyan Chen, Haoyang Lu, Yi Yang, Yufeng Yue
Area: Object Pose Estimation
Datasets: 13-object dynamic interaction test
Metrics: translation error / rotation error / dynamic tracking robustness
Affiliations: Beijing Institute of Technology
DynamicPose: Real-time and Robust 6D Object Pose Tracking for Fast-Moving Cameras and Objects
DynamicPose uses VIO to remove ROI and rotation changes caused by camera ego-motion, and 2D box centers with median depth to compensate for large object translations. World-frame Kalman covariance adaptively generates 1–27 rotation hypotheses, hierarchically filtered by the FoundationPose refiner and fed back into the next frame. High-speed Isaac Sim trajectories of 13 YCB objects yield mean ADD-S / ADD AUC of 88.9 / 75.8.
Tingbang Liang, Yixin Zeng, Jiatong Xie, Boyu Zhou
Area: Object Pose Estimation
Datasets: YCB-Video objects in Isaac Sim / custom real-world mobile-manipulator sequence
Metrics: ADD AUC at 0.1 m / ADD-S AUC at 0.1 m / tracking frequency
Affiliations: Sun Yat-Sen University / Xi’an Jiaotong University / Southern University of Science and Technology / Differential Robotics
EgoGaussian: Dynamic Scene Understanding from Egocentric Video with 3D Gaussian Splatting
EgoGaussian initializes background and object Gaussians from static segments, alternates rigid-object pose optimization and Gaussian densification during dynamic segments, and jointly fine-tunes an egocentric 4D scene representation with per-frame object poses.
Daiwei Zhang, Gengyan Li, Jiajie Li, Mickaël Bressieux, Otmar Hilliges, Marc Pollefeys, Luc Van Gool, Xi Wang
Area: 3D Representations / Reconstruction
Datasets: HOI4D / EPIC-KITCHENS
Metrics: PSNR / SSIM / LPIPS
Affiliations: ETH Zürich / Google / Microsoft / KU Leuven / INSAIT, Sofia
Environment-Agnostic Pose: Generating Environment-independent Object Representations for 6D Pose Estimation
EA6D uses a VAE and an Environment Decoupling Diffusion Model to generate object representations less sensitive to backgrounds, lighting, and occlusion, then decodes rotation and translation. It targets the real-world domain gap using synthetic-only training.
Shaobo Zhang, Yuhang Huang, Wanqing Zhao, Wei Zhao, Ziyu Guan, Jinye Peng
Area: Object Pose Estimation
Datasets: LINEMOD / Occlusion LINEMOD / YCB-Video / DiverseScenes
Metrics: ADD(-S) / 2D Projection / rotation error / translation error
Affiliations: Northwest University / National University of Defense Technology / Xidian University
6-DoF Object Tracking with Event-based Optical Flow and Frames
Local event triplets generate per-event 2D optical flow. A pixel-depth-dependent interaction matrix recovers object 6D velocity in a Kalman filter, while a quaternion UKF fuses 5 Hz absolute DOPE poses. The method substantially outperforms frame-based flow and RGB-D trackers on high-speed synthetic trajectories of four YCB objects. Real dual-camera sequences lack ground truth, and all experiments are offline.
Zhichao Li, Arren Glover, Chiara Bartolozzi, Lorenzo Natale
Area: Object Pose Estimation
Datasets: custom Unreal Engine YCB sequences / custom real event / RGB-D sequences
Metrics: position RMSE / rotation RMSE
Affiliations: Istituto Italiano di Tecnologia / University of Genoa
Multi-Layer Feature Exchange Transformer for Multi-View 6D Object Pose Estimation in Robot Bin Picking
The method adds a multilayer Feature Exchange Transformer to EpiSurfEmb. Self-attention, epipolar cross-attention, and coarse-to-fine scheduling exchange multiview features earlier than final pose fusion. The abstract reports improvements over the baseline and multiview SOTA on ROBI without disclosing numerical results.
Momen Khalil, Vincent Dietrich, Slobodan Ilic
Area: Object Pose Estimation
Datasets: ROBI
Metrics: BOP Average Recall / multi-view 6D pose accuracy
Affiliations: Technical University of Munich / Siemens AG
GCE-Pose: Global Context Enhancement for Category-level Object Pose Estimation
GCE-Pose first reconstructs global semantic geometry from partially visible instances, then fuses reconstruction context with observed features to predict NOCS and category-level pose.
Weihang Li, Hongli Xu, Junwen Huang, Hyunjun Jung, Peter K. T. Yu, Nassir Navab, Benjamin Busam
Area: Object Pose Estimation
Datasets: HouseCat6D / NOCS-REAL275
Metrics: 5° 2cm / 5° 5cm / 10° 2cm / 10° 5cm / 3D IoU
Affiliations: Technical University of Munich / Munich Center for Machine Learning / XYZ Robotics
Generalizable Single-View Object Pose Estimation by Two-Side Generating and Matching
Gen2SM estimates unseen-object relative orientation from one reference and one query. Assuming known reference pose and gravity alignment, Zero-1-to-3 generates 64 intermediate upper-hemisphere views from the reference. Query conditioning explains their noisy latents, and Monte-Carlo denoising residuals form a two-sided pose score. Grid search and SGD refinement recover elevation / azimuth. NAVI 15° / 30° rotation accuracy is 55.32 / 82.14; GSO gives 58.26 / 72.61. Translation, scale, and independent roll are not estimated. Default search and refinement take around 6.5 s + 1.3 s on one A100.
Yujing Sun, Caiyi Sun, Yuan Liu, Yuexin Ma, Siu Ming Yiu
Area: Object Pose Estimation
Datasets: NAVI / Google Scanned Objects (GSO)
Metrics: rotation accuracy within 15 degrees / rotation accuracy within 30 degrees / large-viewpoint rotation accuracy / inference runtime
Affiliations: The University of Hong Kong / ShanghaiTech University
GIVEPose: Gradual Intra-class Variation Elimination for RGB-based Category-Level Object Pose Estimation
GIVEPose represents category-consensus shape with an IVFC coordinate map complementary to NOCS. DCAE progressively suppresses within-instance variation, strengthening RGB-only category-level pose fitting.
Ziqin Huang, Gu Wang, Chenyangguang Zhang, Ruida Zhang, Xiu Li, Xiangyang Ji
Area: Object Pose Estimation
Datasets: NOCS-REAL275 / CAMERA25 / Wild6D
Metrics: N IoU / 10° 0.2d / 10° 0.5d / 5° 5cm
Affiliations: Tsinghua University / Pengcheng Laboratory
Leveraging Global Stereo Consistency for Category-Level Shape and 6D Pose Estimation from Stereo Images
The method represents shared shape within a category as an occupancy field and uses global stereo consistency to constrain coupled shape, scale, and pose optimization, targeting materials with unreliable depth sensing.
Junning Qiu, Minglei Lu, Fei Wang, Yu Guo, Yonggen Ling
Area: Object Pose Estimation
Datasets: TOD / S-RGBD
Metrics: 3D_avg / 3D IoU / 5° 5cm / grasp success
Affiliations: Xi’an Jiaotong University / Tencent Robotics X
GS-Pose: Generalizable Segmentation-Based 6D Object Pose Estimation With 3D Gaussian Splatting
GS-Pose builds semantic, rotation-aware, and 3D Gaussian representations from a few posed RGB reference views of a new object. These support built-in object localization, template-retrieval initialization, and 3DGS render-and-compare refinement.
Dingding Cai, Janne Heikkilä, Esa Rahtu
Area: Object Pose Estimation
Datasets: LINEMOD / OnePose-LowTexture
Metrics: ADD(S)@0.1d / Proj@5pix / mAP@[0.5:0.95]
Affiliations: Tampere University / University of Oulu
Generalizable Zero-Shot Object Pose Estimation for Bin-Picking
GZS6D-BP renders CAD templates in real time for coarse poses, then refines them with camera-view-aligned semantic point-pair features, without retraining for new parts. Its public BOP coarse+refine+teaser submission achieves AR 0.717 across seven Classic-Core datasets, averaging 6.516 s/image.
Zijiang Zhang, Huimin Lu, Jintong Cai, Tohru Kamiya, Seiichi Serikawa
Area: Object Pose Estimation
Datasets: BOP core datasets (7) / Real-world bin-picking system
Metrics: BOP Average Recall / runtime / robotic grasping precision
Affiliations: Kyushu Institute of Technology / Nantong University
HOGSA: Bimanual Hand-Object Interaction Understanding with 3D Gaussian Splatting Based Data Augmentation
HOGSA creates high-quality bimanual hand–object training samples with mesh-based 3D Gaussian Splatting and super-resolution rendering. Optimizing bimanual grasp poses expands viewpoint and pose coverage to improve interaction-pose baselines.
Wentian Qu, Jiahe Li, Jian Cheng, Jian Shi, Chenyu Meng, Cuixia Ma, Hongan Wang, Xiaoming Deng, Yinda Zhang
Area: Object Pose Estimation
Datasets: ARCTIC / H2O
Metrics: CDev / MRRPE / MPJPE / AAE / success rate / PSNR / SSIM / LPIPS
Affiliations: Institute of Automation, Chinese Academy of Sciences / University of Chinese Academy of Sciences / Google
iFusion: Inverting Diffusion for Pose-Free Reconstruction from Sparse Views
iFusion inverts Zero123 novel-view diffusion into an SE(3) view-pose optimizer between two object images with unknown camera poses. Self-estimated poses then guide LoRA fine-tuning and subsequent NeRF / 3DGS reconstruction geometrically.
Chin-Hsuan Wu, Yen-Chun Chen, Bolivar Solarte, Lu Yuan, Min Sun
Area: 3D Representations / Reconstruction
Datasets: Google Scanned Objects / OmniObject3D
Metrics: rotation error / translation error / 5° pose recall / PSNR / SSIM / LPIPS / Chamfer Distance / Volumetric IoU
Affiliations: National Tsing Hua University / Microsoft / Amazon
iG-6DoF: Model-free 6DoF Pose Estimation for Unseen Object via Iterative 3D Gaussian Splatting
iG-6DoF constructs initial poses with multiscale rotation-equivariant features, then refines unknown-object poses through iterative render-and-compare using object 3D Gaussian Splatting.
Tuo Cao, Fei Luo, Jiongming Qin, Yu Jiang, Yusen Wang, Chunxia Xiao
Area: Object Pose Estimation
Datasets: LINEMOD / OnePose-LowTexture / GenMOP / BOP LM
Metrics: ADD(S)-0.1d / Projection 2D / BOP AR_VSD / MSSD / MSPD
Affiliations: Wuhan University
InterTrack: Tracking Human Object Interaction without Object Templates
InterTrack uses CorrAE to establish SMPL correspondences for per-frame human reconstructions and TOPNet to predict occluded-object rotations from temporal and human context. It then jointly optimizes canonical object shape, per-frame transforms, and human–object contact.
Xianghui Xie, Jan Eric Lenssen, Gerard Pons-Moll
Area: Object Pose Estimation
Datasets: ProciGen-Video / BEHAVE / InterCap
Metrics: F-score@1cm / Chamfer Distance / correspondence time
Affiliations: University of Tübingen / Tübingen AI Center / Max Planck Institute for Informatics
JENGA: Object Selection and Pose Estimation for Robotic Grasping from a Stack
JENGA ranks YOLOv8 + ZebraPose hypotheses by object height along IMU gravity and filters occluded blocks using the area ratio of observed modal masks to pose-rendered amodal masks. Its custom synthetic top-1 precision-mAP is 0.79, excluding recall. Real Yaskawa construction-site deployment is shown only in video, without grasp-success measurements.
Sai Srinivas Jeevanandam, Sandeep Inuganti, Shreedhar Govil, Didier Stricker, Jason Rambach
Area: Object Pose Estimation
Datasets: JENGA synthetic stacked-brick dataset / custom real construction-site deployment
Metrics: AP over ADD-S thresholds / AP over MSSD thresholds / mean precision-AP
Affiliations: German Research Center for Artificial Intelligence (DFKI) / RPTU Kaiserslautern
KeyPose: Category-Level 6D Object Pose Estimation with Self-Adaptive Keypoints
KeyPose generates keypoints adaptively from object geometry, aggregates keypoint–keypoint and keypoint–point-cloud graphs, and regresses canonical keypoints through NOCS voxelization to address within-category shape variation structurally.
Sheng Yu, Di-Hua Zhai, Yuanqing Xia
Area: Object Pose Estimation
Datasets: CAMERA25 / NOCS-REAL275 / HouseCat6D
Metrics: 3D IoU / 5° 2cm / 5° 5cm / 10° 2cm / 10° 5cm
Affiliations: Beijing Institute of Technology / Zhongyuan University of Technology
Visuo-Tactile Object Pose Estimation for a Multi-Finger Robot Hand with Low-Resolution In-Hand Tactile Sensing
The method combines visual pose, hand proprioception, binary contact at each finger link, and nonpenetration priors in a factor graph. Robust costs optimize held-object pose under severe hand occlusion or visual noise.
Lukas Mack, Felix Grüninger, Benjamin A. Richardson, Regine Lendway, Katherine J. Kuchenbecker, Joerg Stueckler
Area: Object Pose Estimation
Datasets: 17 YCB objects in simulation / ISyHand real grasp sequences
Metrics: ADD-S / runtime / update rate
Affiliations: University of Augsburg / Max Planck Institute for Intelligent Systems
Improving 6D Object Pose Estimation of Metallic Household and Industry Objects
The paper releases IGD, a BOP-compatible synthetic dataset of 60 metal cans, household objects, and industrial parts. One GDRNPP extension predicts curvature-keypoint heatmaps and fuses them with BAM; another learns difference maps that suppress metallic / specular effects before a second GeoHead pass. Both separately improve synthetic-test AR, without joint training or real-image validation.
Thomas Pöllabauer, Michael Gasser, Tristan Wirth, Sarah Berkei, Volker Knauthe, Arjan Kuijper
Area: Object Pose Estimation
Datasets: IGD metallic object pose dataset
Metrics: BOP Average Recall / ARVSD / ARMSSD / ARMSPD / AD(0.1) / rotation symmetry error / translation symmetry error
Affiliations: Technical University Darmstadt / Fraunhofer IGD / Threedy GmbH
MixRI: Mixing Features of Reference Images for Novel Object Pose Estimation
The method directly fuses features and occlusion states for the same 3D point across reference views, then recovers pose using RANSAC-PnP. It aims to reduce reference counts, memory, and inference costs for BOP novel-object estimation.
Xinhang Liu, Jiawei Shi, Zheng Dang, Yuchao Dai
Area: Object Pose Estimation
Datasets: BOP core datasets / GSO-Datasets
Metrics: BOP AR / inference speed / network parameters / reference image count
Affiliations: Northwestern Polytechnical University / Shaanxi Key Laboratory of Information Acquisition and Processing / EPFL CVLab
MK-Pose: Category-Level Object Pose Estimation via Multimodal-Based Keypoint Learning
MK-Pose fuses frozen DINOv2 RGB features, PointNet++ point cloud features, and CLIP text features from GPT-4o shape descriptions. Soft heatmaps and a GAT discover 96 3D keypoints without manually specified keypoint coordinates. KNN local geometry and global context are aggregated, and NOCS, reconstruction, and finite / continuous-symmetry losses jointly supervise category-level pose and size.
Yifan Yang, Peili Song, Enfan Lan, Dong Liu, Jingtai Liu
Area: Object Pose Estimation
Datasets: CAMERA25 / REAL275 / HouseCat6D
Metrics: 3D IoU25 / IoU50 / IoU75 / 5°/2cm AP / 10°/2cm AP / 10°/5cm AP
Affiliations: Nankai University
MonoDiff9D: Monocular Category-Level 9D Object Pose Estimation via Diffusion Model
MonoDiff9D reconstructs a coarse point cloud from one RGB image using DINOv2. Conditioned on global image / point cloud features and timestep encodings, a Transformer denoiser generates rotation, translation, and scale from Gaussian noise.
Jian Liu, Wei Sun, Hui Yang, Jin Zheng, Zichen Geng, Hossein Rahmani, Ajmal Mian
Area: Object Pose Estimation
Datasets: CAMERA25 / REAL275 / Wild6D
Metrics: 3D IoU / n° mcm / mAP
Affiliations: Hunan University / Central South University / Lancaster University / The University of Western Australia
Occlusion-Aware 6D Pose Estimation with Depth-Guided Graph Encoding and Cross-Semantic Fusion for Robotic Grasping
An ICRA 2025 Best Conference Paper Finalist. A depth-guided GNN models relationships between nonadjacent surface points under occlusion, adaptively fusing masks and binary codes to strengthen RGB-D 2D–3D correspondences. The abstract reports relative gains of 1.2% / 1.9% on LM-O / YCB-V and up to 30% for individual objects.
Jingyang Liu, Zhenyu Lu, Lu Chen, Jing Yang, Chenguang Yang
Area: Object Pose Estimation
Datasets: LM-Occlusion / YCB-Video / Real-world robotic grasping test
Metrics: ADD(-S) / ADD(-S) AUC
Affiliations: Shanxi University / Taiyuan University of Science and Technology / Bristol Robotics Laboratory / University of Liverpool
OmniPose6D: Towards Short-Term Object Pose Tracking in Dynamic Scenes from Monocular RGB
OmniPose6D generates 40,000 dynamic synthetic sequences of 100 frames each from 4,000+ OmniObject3D meshes. Its baseline samples CoTracker trajectories within the first-frame mask, trains uncertainty-aware refinement with five pixel-error levels, and selects co-visible points. COLMAP recovers first-frame-relative, up-to-scale object motion in 8-frame windows. Evaluation requires 7-DoF ground-truth alignment; each window takes around 10 seconds on an RTX 3090.
Yunzhi Lin, Yipu Zhao, Fu-Jen Chu, Xingyu Chen, Weiyao Wang, Hao Tang, Patricio A. Vela, Matt Feiszli, Kevin Liang
Area: Object Pose Estimation
Datasets: OmniPose6D / HO3D v3 / YCBInEOAT / DexYCB
Metrics: KITTI translation error threshold AP / KITTI rotation error threshold AP / ATE threshold AP after 7-DoF alignment / RPE translation / rotation threshold AP
Affiliations: Meta / Georgia Institute of Technology
OmniShape: Zero-Shot Multi-Hypothesis Shape and Pose Estimation in the Real World
OmniShape uses two conditional diffusion models to sample partial NORF point clouds and triplanar shape completions, respectively. Depth registration then places multiple pose–shape hypotheses into the real scene.
Katherine Liu, Sergey Zakharov, Dian Chen, Takuya Ikeda, Greg Shakhnarovich, Adrien Gaidon, Rares Ambrus
Area: 3D Representations / Reconstruction
Datasets: Ocrtoc3D / Pix3D / TYO-L / NOCS REAL275 / HOPE / TUM RGB-D
Metrics: Chamfer-L1 / F1 score / best-of-N hypothesis quality
Affiliations: Toyota Research Institute / Woven by Toyota / Toyota Technological Institute at Chicago
ONDA-Pose: Occlusion-Aware Neural Domain Adaptation for Self-Supervised 6D Object Pose Estimation
ONDA-Pose first fits CAD-like radiance fields from unlabeled real images. A global pose refiner generates pseudo-labels, which are combined with rendered synthetic images for occlusion-aware self-supervised domain adaptation.
Tao Tan, Qiulei Dong
Area: Object Pose Estimation
Datasets: LINEMOD / Occluded LINEMOD / HomebrewedDB
Metrics: ADD(-S) / ADD-S AUC / 2D projection error
Affiliations: University of Chinese Academy of Sciences / CASIA
One2Any: One-Reference 6D Pose Estimation for Any Object
One2Any encodes a single reference RGB-D image into ROPE and decodes Reference Object Coordinates for a query, replacing multiview retrieval and CAD alignment with a continuous relative-coordinate representation.
Mengya Liu, Siyuan Li, Ajad Chhatkuli, Prune Truong, Luc Van Gool, Federico Tombari
Area: Object Pose Estimation
Datasets: REAL275 / Toyota-Light / YCB-Video / LINEMOD
Metrics: ADD-S AUC / ADD AUC / BOP AR / tracking AUC / inference time
Affiliations: ETH Zurich / INSAIT / Sofia University / Google / Technical University of Munich
PACA: Perspective-Aware Cross-Attention Representation for Zero-Shot Scene Rearrangement
PACA unifies Stable Diffusion generation, object-region extraction, and feature encoding into word-conditioned cross-attention representations. Targets are generated directly; real RGB undergoes DDIM inversion / reconstruction before object representations are read along the same denoising timeline. ICP, Marigold relative depth, and RGB-D alignment compute rearrangement transforms. Mean matching accuracy on 450 real–generated pairs is 86.72% versus DALL·E-Bot's 51.35%. Across 60 real robot trials, mean success in 3-DoF / 6-DoF settings is 68.73% / 65.56%. The text explicitly defines object transforms as (x,y,z,θ), and success measures relative layouts rather than pose ground truth. This is a generative-goal robotic extension, not a full SE(3) object-pose benchmark.
Shutong Jin, Ruiyu Wang, Kuangyi Chen, Florian T. Pokorny
Area: Robot Perception & Manipulation
Datasets: PACA Dining Scene / PACA Fruit Scene / PACA Office Scene / 450 real–generated image pairs / 60 real-robot rearrangement trials
Metrics: object matching accuracy / robot execution success rate / Inception Score / CLIP Score / human goal-satisfaction rating
Affiliations: KTH Royal Institute of Technology / Graz University of Technology
PartPose: Attentive 6D Pose Estimation by Focusing on Graspable Parts of Multi-Part Deformable Objects
PartPose targets mixed rigid–deformable objects such as drink pouches, wired appliances, and objects with cords. DINOv2 transfers reference keypoints, SAM segments graspable rigid ROIs, Bayesian optimization selects ROI geometry, and FoundationPose performs local render-and-compare. Across nine objects, overall translation / rotation success is 98.2% / 96.4%, with real grasp success of 87.2%. CAD, RGB-D, manual reference points, and visible rigid parts are required; fully deformable targets are unsupported. The formal publication is RA-L 2025, with ICRA 2026 only hosting its journal-paper presentation.
Ryo Okumura, Tadahiro Taniguchi
Area: Object Pose Estimation
Datasets: PartPose multi-part deformable object dataset
Metrics: translation success rate at 30 mm / rotation success rate at 0.3 rad / robotic picking success rate / translation error / rotation error / inference latency
Affiliations: Panasonic Holdings Corporation / Kyoto University / Ritsumeikan University
Category-Level 6D Object Pose Estimation in Agricultural Settings Using a Lattice-Deformation Framework and Diffusion-Augmented Synthetic Data
PLANTPose explicitly represents category-instance deformation with a 24D cubic B-spline lattice and removes its rigid component with Umeyama alignment. Blender physics scenes and depth-conditioned Stable Diffusion generate training images. A small ViT jointly predicts rotation, deformation, and projections from RGB, while PnP recovers camera-frame translation.
Marios Glytsos, Panagiotis P. Filntisis, George Retsinas, Petros Maragos
Area: Object Pose Estimation
Datasets: PLANTPose banana benchmark / PLANTPose synthetic training set
Metrics: Chamfer distance / mean / median rotation error / mean / median translation error / deformation error
Affiliations: Robotics Institute, Athena Research and Innovation Center / National Technical University of Athens
Pos3R: 6D Pose Estimation for Unseen Objects Made Easy
Pos3R connects 3D-consistent features from a 3D reconstruction foundation model to template selection and correspondence estimation, handling unknown objects without pose-specific training.
Weijian Deng, Dylan Campbell, Chunyi Sun, Jiahao Zhang, Shubham Kanitkar, Matthew E. Shaffer, Stephen Gould
Area: Object Pose Estimation
Datasets: BOP: YCB-V / LM-O / T-LESS / TUD-L / IC-BIN / ITODD / HB
Metrics: BOP AR / AR_VSD / AR_MSSD / AR_MSPD
Affiliations: Australian National University / RIOS Intelligent Machines
QORT-Former: Query-optimized Real-time Transformer for Understanding Two Hands Manipulating Objects
QORT-Former uses separate left-hand / right-hand / object queries, a query proposal loss, and contact features to jointly estimate both hands and object 3D poses from one RGB image in real time, while supporting interaction recognition.
Elkhan Ismayilzada, MD Khalequzzaman Chowdhury Sayem, Yihalem Yimolal Tiruneh, Mubarrat Tajoar Chowdhury, Muhammadjon Boboev, Seungryul Baek
Area: Object Pose Estimation
Datasets: H2O / First-Person Hand Action (FPHA)
Metrics: mean end-point error / Top-1 accuracy / FPS
Affiliations: Ulsan National Institute of Science and Technology / Michigan State University
R²-Art: Category-Level Articulation Pose Estimation from Single RGB Image via Cascade Render Strategy
R²-Art estimates root pose and joint states in parallel from a single RGB image. Joint-centric priors transform / deform individual parts, and cascaded rendering losses progressively correct category-level articulated-part poses.
Li Zhang, Haonan Jiang, Yukang Huo, Yan Zhong, Jianan Wang, Xue Wang, Rujing Wang, Liu Liu
Area: Object Pose Estimation
Datasets: ArtImage / ReArtMix / RobotArm
Metrics: ADD / AUC / 3D IoU
Affiliations: Hefei Institutes of Physical Science, Chinese Academy of Sciences / University of Science and Technology of China / Astribot / Zhejiang University of Technology / China Agricultural University / Peking University / Hefei University of Technology
RAG-6DPose: Retrieval-Augmented 6D Pose Estimation via Leveraging CAD as Knowledge Base
RAG-6DPose back-projects DINOv2 appearance features from multiview CAD renderings to 3D points, forming per-object memory with coordinates and colors. Query RGB uses ReSPC self-attention, PointNet, and cross-attention to retrieve relevant CAD points. SurfEmb-style InfoNCE learns dense 2D–3D correspondences, and PnP-RANSAC with optional ICP produces pose.
Kuanning Wang, Yuqian Fu, Tianyu Wang, Yanwei Fu, Longfei Liang, Yu-Gang Jiang, Xiangyang Xue
Area: Object Pose Estimation
Datasets: LM-O / IC-BIN / TUD-L / YCB-V / HB
Metrics: BOP Average Recall / ARVSD / ARMSSD / ARMSPD / robot task success rate
Affiliations: Fudan University / INSAIT, Sofia University / NeuhHelium Co., Ltd.
Shape-Biased Texture Agnostic Representations for Improved Textureless and Metallic Object Detection and 6D Pose Estimation
Without changing GDR-Net, Pix2Pose, or detector architectures, the paper randomly samples from 1,226 textures and UV-maps them onto objects during BlenderProc PBR rendering, reducing CNN dependence on fixed colors and materials. The full shape-biased detection–pose pipeline improves GDR-Net / Pix2Pose BOP AR from 56.29 / 46.68 to 58.32 / 50.13 on T-LESS and from 10.70 / 12.90 to 11.10 / 15.30 on ITODD. Gaussian noise, shot noise, and brightness robustness improve, while motion / Gaussian blur barely improves. Gains depend on detection and online augmentation; CAD + shape bias reaches 62.95 AR versus reconstructed meshes + shape bias at 46.57, so random textures do not replace accurate geometry.
Peter Hönig, Stefan Thalhammer, Jean-Baptiste Weibel, Matthias Hirschmanner, Markus Vincze
Area: Object Pose Estimation
Datasets: T-LESS / ITODD
Metrics: BOP Average Recall / MSPD Average Recall / MSSD Average Recall / VSD Average Recall / COCO detection mAP / image-perturbation robustness
Affiliations: TU Wien / UAS Technikum Vienna
RayPose: Ray Bundling Diffusion for Template Views in Unseen 6D Object Pose Estimation
Template-based pose estimation is formulated as object-centered camera-ray alignment. A diffusion Transformer conditioned on multiview templates with known poses denoises structured rotation / translation maps for coarse-to-fine inference.
Junwen Huang, Shishir Reddy Vutukur, Peter KT Yu, Nassir Navab, Slobodan Ilic, Benjamin Busam
Area: Object Pose Estimation
Datasets: LINEMOD / Occlusion LINEMOD / YCB-Video / T-LESS
Metrics: ADD(-S) Recall / BOP AR / rotation error / translation error
Affiliations: Technical University of Munich / Munich Center for Machine Learning / XYZ Robotics
RCGNet: RGB-based Category-Level 6D Object Pose Estimation with Geometric Guidance
RCGNet uses frozen DINOv2 to predict and fuse appearance and geometry from RGB crops. Ground-truth NOCS point cloud features guide geometry during training; category-prior deformation, a scale branch, and RANSAC-PnP recover category-level pose. Inference uses no depth, but category point clouds, masks, and training-time geometric supervision remain necessary.
Sheng Yu, Di-Hua Zhai, Yuanqing Xia
Area: Object Pose Estimation
Datasets: CAMERA25 / REAL275 / Wild6D
Metrics: 3D IoU50 / 3D IoU75 / 10 cm mAP / 10° mAP / 10°10 cm mAP
Affiliations: Beijing Institute of Technology / Zhongyuan University of Technology
Comparative Evaluation of 3D Reconstruction Methods for Object Pose Estimation
The benchmark captures 397–505 calibrated high-resolution views for each of 21 physical YCB-V objects, comparing BOP AR when COLMAP, RealityCapture, Nerfacto, VolSDF, MonoSDF, BakedSDF, and other reconstructions feed MegaPose / FoundationPose. Geometry completeness / accuracy, PSNR, and downstream pose rankings differ. Even the best reconstructions typically trail CAD by around 10%–20%; RealityCapture and Nerfacto offer favorable accuracy–time tradeoffs, and most methods approach full-data performance with 75–100 images. Small, reflective, finely textured, and physically changed objects are hardest. Benchmark registration still initializes ICP with CAD and manual coarse alignment, so this is not complete CAD-free onboarding.
Varun Burde, Assia Benbihi, Pavel Burget, Torsten Sattler
Area: 3D Representations / Reconstruction
Datasets: YCB-Video / Reconstruction Pose Benchmark for 21 YCB-V objects
Metrics: BOP Average Recall / VSD recall / MSSD recall / MSPD recall / geometry completeness / geometry accuracy / PSNR / reconstruction runtime / image-count efficiency
Affiliations: Czech Institute of Informatics, Robotics and Cybernetics, Czech Technical University in Prague
Recurrent Feature Mining and Keypoint Mixup Padding for Category-Agnostic Pose Estimation
For category-agnostic keypoint pose estimation, recursive deformable attention extracts fine-grained support / query structure. Mixup keypoint padding accommodates variable keypoint counts.
Junjie Chen, Weilong Chen, Yifan Zuo, Yuming Fang
Area: Object Pose Estimation
Datasets: MP-100
Metrics: mPCK / PCK@0.05 / 1-shot / 5-shot accuracy
Affiliations: Jiangxi University of Finance and Economics
RefPose: Leveraging Reference Geometric Correspondences for Accurate 6D Pose Estimation of Unseen Objects
RefPose initializes pose from reference geometric correspondences and uses correlation-volume-guided attention for render-and-compare refinement, reducing dependence on fixed object models for unknown objects.
Jaeguk Kim, Jaewoo Park, Keuntek Lee, Nam Ik Cho
Area: Object Pose Estimation
Datasets: BOP: YCB-V / LM-O / T-LESS / TUD-L / IC-BIN / ITODD / HB
Metrics: BOP AR / AR_VSD / AR_MSSD / AR_MSPD / runtime
Affiliations: Seoul National University / INMC / IPAI
RGBTrack: Fast, Robust Depth-Free 6D Pose Estimation and Tracking
RGBTrack replaces first-frame median depth with binary search matching rendered and observed mask areas using metric CAD. Later frames zero FoundationPose's point cloud channels and refine from RGB and the previous pose. XMem, Kalman filtering, and a state machine sample 20 recovery hypotheses when centers / areas become anomalous. Metric recovery for CAD of unknown scale still requires true first-frame depth.
Teng Guo, Jingjin Yu
Area: Object Pose Estimation
Datasets: YCBInEOAT / ClearPose
Metrics: BOP Average Recall / translation error / rotation error / tracking FPS
Affiliations: Rutgers, The State University of New Jersey
RoMa v2: Harder Better Faster Denser Feature Matching
RoMa v2 is a dense feature matcher for downstream visual localization, SfM, reconstruction, and pose estimation. DINOv3, a multiview Transformer, lightweight refinement, and predicted covariance improve matching accuracy, speed, and generalization in difficult scenes.
Johan Edstedt, David Nordstrom, Yushan Zhang, Georg Bokman, Jonathan Astermark, Viktor Larsson, Anders Heyden, Fredrik Kahl, Marten Wadenback, Michael Felsberg
Area: Feature Matching / Correspondences
Datasets: MegaDepth-1500 / ScanNet-1500 / TA-WB / ScanNet++ v2 / FlyingThings3D / AerialMegaDepth / MapFree / WxBS / SatAst / HyperSim
Metrics: AUC@5/10/20 / EPE / PCK@1/3/5px / mAA@10px / AUC@10px / throughput / memory
Affiliations: Linkoping University / Chalmers University of Technology / University of Amsterdam / Centre for Mathematical Sciences, Lund University
Scaling 3D Compositional Models for Robust Classification and Pose Estimation
Within-class / between-class decomposition and dynamic hard-negative focusing scale compositional vertex features in Neural Mesh Models to 188 classes. Classification and pose robustness are evaluated under occlusion, weather, synthetic-data, and unseen-category OOD shifts.
Xiaoding Yuan, Guofeng Zhang, Prakhar Kaushik, Artur Jesslen, Adam Kortylewski, Alan Yuille
Area: Object Pose Estimation
Datasets: ImageNet3D+ / OOD-CV
Metrics: classification accuracy / pose estimation accuracy / training time / OOD robustness
Affiliations: Johns Hopkins University / University of Freiburg / Max Planck Institute for Informatics
SCFlow2: Plug-and-Play Object Pose Refiner with Shape-Constraint Scene Flow
SCFlow2 uses RGB-D 3D scene flow as a shape constraint for recurrent matching. Its pose refiner can improve different initial estimators as a postprocessing module without retraining for new objects.
Qingyuan Wang, Rui Song, Jiaojiao Li, Kerui Cheng, David Ferstl, Yinlin Hu
Area: Object Pose Estimation
Datasets: BOP LM / YCB-Video / Objaverse / Google Scanned Objects / ShapeNet
Metrics: BOP AR / ADD(-S) / ADD-S AUC / refinement gain
Affiliations: Xidian University / Taiyuan University of Technology / Magic Leap
SDFit: 3D Object Pose and Shape by Fitting a Morphable SDF to a Single Image
The method retrieves and fits a deformable SDF from a single image, jointly optimizing shape and pose in a render-and-compare loop. Foundation features initialize 2D–3D alignment, supporting generalization without retraining to occlusion, unusual poses, and in-the-wild images.
Dimitrije Antić, Georgios Paschalidis, Shashank Tripathi, Theo Gevers, Sai Kumar Dwivedi, Dimitrios Tzionas
Area: Object Pose Estimation
Datasets: Pix3D / Pascal3D+ / COMIC
Metrics: 3D Chamfer distance / 2D IoU / pose error / shape reconstruction metrics
Affiliations: University of Amsterdam / Max Planck Institute for Intelligent Systems
Semantic Cross-Pose Correspondence from a Single Example
Semantic Cross-Pose automatically localizes interaction regions in demonstrations, learning soft inter-object correspondences from DINO-ViT semantics and local geometry. Differentiable SVD transfers the bottleneck relative pose to unseen object pairs.
Denis Hadjivelichkov, Sicelukwanda Zwane, Marc Peter Deisenroth, Lourdes Agapito, Dimitrios Kanoulas
Area: Robot Perception & Manipulation
Datasets: GSOD / PyBullet interaction pairs / Real RGB-D interaction demonstrations / Human preference survey
Metrics: task success rate / survey score / cross-pose consistency
Affiliations: University College London / Archimedes / Athena Research Center
SMORE: Simultaneous Map and Object REconstruction
SMORE decomposes dynamic urban LiDAR scenes into background and rigid objects, alternating optimization of neural surfaces, ego poses, and object poses. Extended continuous-time deskewing corrects rolling-shutter LiDAR distortions of moving objects.
Nathaniel Chodosh, Anish Madan, Simon Lucey, Deva Ramanan
Area: Object Pose Estimation
Datasets: nuScenes / Argoverse 2
Metrics: Average Translation Error / nearest-neighbor distance / Chamfer Distance / median depth error / PSNR / SSIM / LPIPS
Affiliations: Villanova University / Carnegie Mellon University / University of Adelaide
Rethinking Correspondence-based Category-Level Object Pose Estimation
SpotPose addresses two central bottlenecks in category-level correspondence pipelines: shape-sensitive / pose-invariant features and outlier removal. Incorrect correspondences are explicitly filtered before fitting.
Huan Ren, Wenfei Yang, Shifeng Zhang, Tianzhu Zhang
Area: Object Pose Estimation
Datasets: NOCS-REAL275 / CAMERA25 / HouseCat6D
Metrics: N IoU / 5° 5cm / 10° 5cm / 3D IoU
Affiliations: University of Science and Technology of China / National Key Laboratory of Deep Space Exploration / Sangfor Technologies
Differentiable Rendering-based Pose Estimation for Surgical Robotic Instruments
From a single endoscopic RGB image, SAM2 mask, dVRK CAD / kinematics, and inaccurate joint angles, the method batch-optimizes 500 initializations in a four-dimensional LookAt space, then jointly refines camera-to-robot extrinsics and three distal joints. Results on 30 real frames improve RCM geometric consistency, without absolute extrinsic ground truth or runtime results. It initializes a markerless tracker rather than providing real-time tracking.
Zekai Liang, Zih-Yun Chiu, Florian Richter, Michael C. Yip
Area: Object Pose Estimation
Datasets: custom real dVRK calibration frames / custom gauze and suture-needle picking trials
Metrics: RCM converging-point distance standard deviation / open-loop picking success rate
Affiliations: University of California San Diego (UCSD)
SurgiPose: Estimating Surgical Tool Kinematics from Monocular Video for Surgical Robot Learning
SurgiPose initializes 324 candidates from a 3×3 mask-center grid and 36 viewing-axis angles, then jointly refines tool 6D pose and three joint angles with joint-conditioned 3D Gaussian Splatting. Subsequent frames use warm-start optimization. ADE on two tasks is 9.7 / 12.0 mm; ACT policies supervised by estimated trajectories achieve 70% / 60% success versus 100% / 80% with ground truth. Speed and quantitative cross-dataset generalization remain unreported.
Juo-Tung Chen, XinHao Chen, Ji Woong Kim, Paul Maria Scheikl, Richard Jaepyeong Cha, Axel Krieger
Area: Object Pose Estimation
Datasets: custom dVRK Si tissue-lifting demonstrations / custom dVRK Si needle-pickup demonstrations / SurgRIPE / custom ex vivo cholecystectomy dataset
Metrics: Average Displacement Error / Final Displacement Error / per-axis trajectory error / policy task success rate / policy execution time
Affiliations: Johns Hopkins University / Optosurgical
UA-Pose: Uncertainty-Aware 6D Object Pose Estimation and Online Object Completion with Partial References
UA-Pose distinguishes observed and unobserved regions through a hybrid object model with uncertainty, allowing incomplete references to support pose confidence, robust estimation, and online shape completion.
Ming-Feng Li, Xin Yang, Fu-En Wang, Hritam Basak, Yuyin Sun, Shreekant Gayaka, Min Sun, Cheng-Hao Kuo
Area: Object Pose Estimation
Datasets: YCB-Video / YCBInEOAT / HO3D
Metrics: ADD-S AUC / ADD AUC / AR / object completion
Affiliations: Carnegie Mellon University / Stony Brook University / National Tsing Hua University / Amazon
Ultra-Precision 6DoF Pose Estimation Using 2-D Interpolated Discrete Fourier Transform
A 2D interpolated DFT estimates 6-DoF pose from the phase and frequency of periodic patterns, with a mathematical mapping from image parameters to pose. It targets controlled industrial measurements at nanometer translation and microradian rotation scales.
Guowei Shi, Zian Mao, Peisen Huang
Area: Object Pose Estimation
Datasets: periodic-pattern simulation / industrial camera measurement setup
Metrics: translation RMSE / rotation RMSE / nanometer accuracy / microradian accuracy
Affiliations: UM-SJTU Joint Institute, Shanghai Jiao Tong University
Uncertainty-Aware Knowledge Distillation for Compact and Efficient 6DoF Pose Estimation
The method estimates epistemic uncertainty from 2D keypoint variance across 4–6 independent teachers, reducing distillation mass for unstable teacher points in unbalanced optimal transport. PFKD reuses the transport plan to align corresponding keypoint feature receptive fields, compressing WDRNet / SPNv2 to 2.3–3.8M parameters while improving LINEMOD and SPEED+ pose results.
Nassim Ali Ousalah, Anis Kacem, Enjie Ghorbel, Emmanuel Koumandakis, Djamila Aouada
Area: Object Pose Estimation
Datasets: LINEMOD / SPEED+
Metrics: ADD(-S)-0.1d / rotation error / translation error / pose error / parameters / FLOPs
Affiliations: University of Luxembourg / ENSI / University of Manouba / Infinite Orbits
Unified Category-Level Object Detection and Pose Estimation from RGB Images using 3D Prototypes
A neural mesh with learned vertex features serves as a category-level 3D prototype. A single-stage RGB framework jointly performs detection, 6D pose estimation, and 9D pose refinement, avoiding a cascade of detection and single-object pose networks.
Tom Fischer, Xiaojie Zhang, Eddy Ilg
Area: Object Pose Estimation
Datasets: REAL275 / CAMERA25
Metrics: 3D IoU / rotation error / translation error / scale error / mAP
Affiliations: Saarland University / University of Technology Nuremberg
Universal Features Guided Zero-Shot Category-Level Object Pose Estimation
Universal Features Pose uses general-purpose 2D features for sparse semantic correspondences and coarse pose, then general-purpose 3D features for dense alignment, shape optimization, and pose refinement, enabling zero-shot category-level estimation without additional fine-tuning.
Wentian Qu, Chenyu Meng, Heng Li, Jian Cheng, Cuixia Ma, Hongan Wang, Xiao Zhou, Xiaoming Deng, Ping Tan
Area: Object Pose Estimation
Datasets: NOCS-REAL275 / Wild6D
Metrics: 3D IoU / 5° 2cm / 5° 5cm / 10° 2cm / 10° 5cm
Affiliations: Institute of Software, Chinese Academy of Sciences / University of Chinese Academy of Sciences / The Hong Kong University of Science and Technology / Aerospace Information Research Institute, Chinese Academy of Sciences
UNOPose: Unseen Object Pose Estimation with an Unposed RGB-D Reference Image
UNOPose reduces reference acquisition to a single unposed RGB-D image, canonicalizing the object with an SE(3)-invariant reference frame and reweighting correspondences by overlap likelihood.
Xingyu Liu, Gu Wang, Ruida Zhang, Chenyangguang Zhang, Federico Tombari, Xiangyang Ji
Area: Object Pose Estimation
Datasets: UNOPose benchmark / BOP: LM-O / TUD-L / YCB-V
Metrics: BOP AR / segmentation mAP / runtime
Affiliations: Tsinghua University / Technical University of Munich / Google
VAPO: Visibility-Aware Keypoint Localization for Efficient 6DoF Object Pose Estimation
VAPO separates keypoint visibility into external occlusion determined by masks and internal self-occlusion approximated by surface normals. Personalized PageRank diffuses binary labels into continuous importance on the object's k-NN graph. The network localizes only the top 256 / 512 keypoints for the current image, followed by Progressive-X / PnP, and can substitute multiview-reconstructed meshes for exact CAD.
Ruyi Lian, Yuewei Lin, Longin Jan Latecki, Haibin Ling
Area: Object Pose Estimation
Datasets: LINEMOD / LINEMOD-Occlusion / YCB-Video
Metrics: ADD(-S) recall at 0.02d / 0.05d / 0.1d / AUC of ADD-S / AUC of ADD(-S) / runtime
Affiliations: South Dakota State University / Brookhaven National Laboratory / Temple University / Stony Brook University
Free-Moving Object Reconstruction and Pose Estimation with Virtual Camera
Virtual Camera Pose jointly reconstructs shape and continuous poses from monocular freely moving-object sequences. It first reduces the optimization space in a virtual camera frame, then uses PnP to return to the real camera frame and performs global refinement.
Haixin Shi, Yinlin Hu, Daniel Koguciuk, Juan-Ting Lin, Mathieu Salzmann, David Ferstl
Area: Object Pose Estimation
Datasets: HO3D / egocentric free-moving object sequences
Metrics: HDRM SE / AUC ATE / relative rotation error / relative translation error
Affiliations: École polytechnique fédérale de Lausanne / Magic Leap
ViTa-Zero: Zero-Shot Visuotactile Object 6D Pose Estimation
ViTa-Zero checks whether visual poses satisfy contact, penetration, and kinematic constraints. If infeasible, it models hand–object interaction as a spring–mass system and optimizes FoundationPose / MegaPose results at test time.
Hongyu Li, James Akl, Srinath Sridhar, Tye Brady, Taskin Padir
Area: Object Pose Estimation
Datasets: Five-object real visuotactile manipulation set
Metrics: ADD-S AUC / ADD AUC / position error / runtime
Affiliations: Amazon Fulfillment Technologies & Robotics / Brown University / Northeastern University
You Only Estimate Once: Unified, One-Stage, Real-Time Category-Level Articulated Object 6D Pose Estimation for Robotic Grasping
YOEO predicts part semantics, centroid offsets, and NPCS in parallel with a single point cloud network. Voting-based clustering separates part instances, followed by NPCS-to-observation registration for real-time articulated-object manipulation.
Jingshun Huang, Haitao Lin, Tianyu Wang, Yanwei Fu, Yu-Gang Jiang, Xiangyang Xue
Area: Object Pose Estimation
Datasets: GAPartNet / Real-world Kinova manipulation set
Metrics: rotation error / translation error / scale error / 3D mIoU / robot success rate
Affiliations: Fudan University / Tencent Robotics X
ZeroBP: Learning Position-Aware Correspondence for Zero-Shot 6D Pose Estimation in Bin-Picking
ZeroBP combines local point cloud features with object-centric direction encoding in cross-attention. Alternating superpoint-to-point pose / position refinement reduces mismatches caused by repetitive local geometry in industrial parts.
Jianqiu Chen, Zikun Zhou, Xin Li, Ye Zheng, Tianpeng Bao, Zhenyu He
Area: Object Pose Estimation
Datasets: Google Scanned Objects synthetic set / ROBI
Metrics: AR of ADD(-S) / ADD(-S)-0.1d
Affiliations: Harbin Institute of Technology, Shenzhen / Pengcheng Laboratory / JD.com / SenseTime Research
ZeroPose: CAD-Prompted Zero-Shot Object 6D Pose Estimation in Cluttered Scenes
For CAD-based RGB-D 6D pose estimation of unseen rigid instances, each target CAD model is rendered offline from multiple views, caching DINOv2 image / patch and geometry features. FastSAM proposals and global visual matching discover all target instances; local patch matching predicts visible views, followed by coarse-to-fine 3D–3D matching and weighted SVD between visible CAD surfaces and scene point clouds. New objects require no retraining or manually supplied masks, but metric CAD, depth, intrinsics, and one-time onboarding remain necessary.
Jianqiu Chen, Zikun Zhou, Mingshan Sun, Rui Zhao, Liwei Wu, Tianpeng Bao, Zhenyu He
Area: Object Pose Estimation
Datasets: Google Scanned Objects / BOP core datasets / LM-O / T-LESS / TUD-L / IC-BIN / ITODD / HomebrewedDB / YCB-Video
Metrics: instance segmentation mAP / BOP AR (VSD/MSSD/MSPD) / ADD-S 2 cm / inference time / onboarding time
Affiliations: Harbin Institute of Technology, Shenzhen / Pengcheng Laboratory / SenseTime Research
6D-Diff: A Keypoint Diffusion Framework for 6D Object Pose Estimation
6D-Diff formulates 2D keypoint localization as conditional reverse diffusion from heatmap distributions to definite coordinates. A 9-component Mixture-of-Cauchy models heavy-tailed, multimodal errors from occlusion and clutter. Mean LM-O ADD(-S) is 79.6; YCB-V ADD(-S) / ADD-S AUC / ADD(-S) AUC is 83.8 / 91.5 / 87.0. Without MoC, LM-O drops to 73.1.
Li Xu, Haoxuan Qu, Yujun Cai, Jun Liu
Area: Object Pose Estimation
Datasets: LM-O / YCB-Video
Metrics: ADD(-S) / AUC of ADD-S / AUC of ADD(-S)
Affiliations: Singapore University of Technology and Design / Nanyang Technological University
Shape From Shading for Robotic Manipulation
The system captures seven static RGB images with two cameras, six low-incidence directional lights, and one top light. Calibrated photometric stereo recovers normals, depth edges, and optional surfaces of small matte-white Lambertian objects. Mean surface errors for ten printed objects range from 0.50–1.74 mm. Known-object CAD alignment is measured by 0.80–5.50 px edge misalignment rather than ground-truth R/t. Unknown-object tests cover only relative planar x/y/theta, with maximum errors of 3.28 mm, 2.40 mm, and 1.28°. It demonstrates active lighting for grasping and pose, without establishing natural-material, single-frame, or standard full-6D benchmark performance.
Arkadeep Narayan Chaudhury, Leonid Keselman, Christopher G. Atkeson
Area: Robot Perception & Manipulation
Datasets: custom controlled-illumination tabletop objects / custom robot suction-grasp trials / custom PVC-card deformation experiment / custom known-object pose and unknown-object tracking tests / DiLiGenT
Metrics: surface-normal angular error / integrated-surface error in millimeters / image-edge pixel misalignment / planar translation error / planar rotation error / normal-distribution dissimilarity / grasp trial outcomes / deformation fit R-squared
Affiliations: Carnegie Mellon University
Instance-Adaptive and Geometric-Aware Keypoint Learning for Category-Level 6D Object Pose Estimation
For category-level RGB-D 6D pose estimation, the method learns instance-adaptive sparse keypoints and uses local / global geometry-aware feature aggregation to establish keypoint-level NOCS correspondences. It avoids category shape priors and improves unseen-instance generalization on REAL275 and CAMERA25.
Xiao Lin, Wenfei Yang, Yuan Gao, Tianzhu Zhang
Area: Object Pose Estimation
Datasets: CAMERA25 / REAL275 / NOCS
Metrics: 3D IoU / 5° 2cm / 5° 5cm / 10° 2cm / 10° 5cm
Affiliations: University of Science and Technology of China / Jianghuai Advance Technology Center
Learning to Estimate 6DoF Pose from Limited Data: A Few-Shot, Generalizable Approach using RGB Images
Cas6D uses a self-supervised ViT for robust few-shot object detection, retains top-K reference pose hypotheses, and corrects large initialization errors with multiscale warped feature volumes and progressively narrowed discrete pose bins.
Panwang Pan, Zhiwen Fan, Brandon Yushan Feng, Peihao Wang, Chenxin Li, Zhangyang Wang
Area: Object Pose Estimation
Datasets: LINEMOD / GenMOP
Metrics: ADD(S)-0.1d / Proj-5 / mAP@[.5:.95]
Affiliations: ByteDance / The University of Texas at Austin / University of Maryland / The Hong Kong Polytechnic University
Category-level Object Detection, Pose Estimation and Reconstruction from Stereo Images
CODERS encodes metric 3D coordinates sampled along stereo frusta into ConvNeXt features. DETR queries jointly predict multi-object detections, 6D poses, sizes, and DeepSDF meshes. TOD bottle/mug 3D50 is 99.5/100, but strict bottle 3D75 is 31.5. SS3D covers 16 categories and 427 instances; code / data are available, while an RTX3090 achieves only 3Hz in the fixed-camera setting.
Chuanrui Zhang, Yonggen Ling, Minglei Lu, Minghan Qin, Haoqian Wang
Area: Object Pose Estimation
Datasets: SS3D / TOD / OmniObject3D
Metrics: 3D IoU precision at 25% / 50% / 75% / 5°2cm / 5°5cm / 10°5cm / 10°10cm / Chamfer distance / inference FPS
Affiliations: Tsinghua University / Tencent Robotics X
ContactArt: Learning 3D Interaction Priors for Category-level Articulated Object and Hand Poses Estimation
ContactArt learns a part-arrangement discriminator and contact diffusion prior from simulated human teleoperation data, then uses both to constrain category-level articulated-object and hand poses in real images.
Zehao Zhu, Jiashun Wang, Yuzhe Qin, Deqing Sun, Varun Jampani, Xiaolong Wang
Area: Object Pose Estimation
Datasets: ContactArt / HOI4D / BMVC articulated-object dataset / RBO
Metrics: 5°5cm / mIoU / rotation error / translation error / hand MPJPE
Affiliations: The University of Texas at Austin / Carnegie Mellon University / UC San Diego / Google Research
Object Pose Estimation via the Aggregation of Diffusion Features
The method freezes Stable Diffusion v1.5 and aggregates intermediate U-Net features for template matching. CWA achieves Acc15 of 98.2 / 85.9 on unseen LM / O-LM and VSD recall of 71.03 on T-LESS with 21k templates. Generalization of unseen-object representations is the benefit; CAD, ground-truth crops, and large template libraries remain necessary, and complete 6D pose is not directly regressed.
Tianfu Wang, Guosheng Hu, Hongguang Wang
Area: Object Pose Estimation
Datasets: LINEMOD / Occlusion LINEMOD / T-LESS
Metrics: Acc15 rotation-and-class accuracy / Recall VSD at 0.3 / trainable parameter count
Affiliations: Shenyang Institute of Automation, Chinese Academy of Sciences / Institutes for Robotics and Intelligent Manufacturing, Chinese Academy of Sciences / University of Chinese Academy of Sciences / Oosto
DiffusionNOCS: Managing Symmetry and Uncertainty in Sim2Real Multi-Modal Category-Level Pose Estimation
DiffusionNOCS generates multiple dense NOCS maps with conditional diffusion, then uses depth–NOCS 3D correspondences and TEASER++ to recover scale and 6D pose. Multimodal conditioning and multiple noise hypotheses address symmetry-induced alternatives, missing depth, and synthetic-only training's transfer to real environments.
Takuya Ikeda, Sergey Zakharov, Tianyi Ko, Muhammad Zubair Irshad, Robert Lee, Katherine Liu, Rares Ambrus, Koichi Nishiwaki
Area: Object Pose Estimation
Datasets: NOCS REAL275 / TYO-L / YCB-V / HOPE / ShapeNet synthetic training set
Metrics: 5°5cm mAP / 10°5cm mAP / 15°5cm mAP / inference time
Affiliations: Woven by Toyota / Toyota Research Institute
Neural Implicit Representation for Building Digital Twins of Unknown Articulated Objects
DigitalTwinArt first reconstructs object-level neural SDFs for two articulation states, then jointly recovers a multijoint digital twin through part segmentation, shared rigid motion, and an explicit correspondence field. In multipart ablations, full-model axis-angle / mixed-unit part-motion / moving-part CD errors are 0.26° / 0.09 / 0.33 mm. Removing shared motion worsens them to 14.42° / 0.28 / 127.08 mm.
Yijia Weng, Bowen Wen, Jonathan Tremblay, Valts Blukis, Dieter Fox, Leonidas Guibas, Stan Birchfield
Area: Object Pose Estimation
Datasets: PARIS Two-Part Object Dataset / Synthetic Multi-Part Object Dataset
Metrics: joint axis angle / position error / part motion error / Chamfer Distance for static / moving / whole shape
Affiliations: NVIDIA / Stanford University
RGB-Based Category-Level Object Pose Estimation via Decoupled Metric Scale Recovery
DMSR decouples correspondence learning from metric-scale recovery for RGB-only category-level pose. Monocular depth / normals provide local geometry, while a separate scale branch prevents scale errors from contaminating PnP, jointly recovering pose and size for unseen category instances.
Jiaxin Wei, Xibin Song, Weizhe Liu, Laurent Kneip, Hongdong Li, Pan Ji
Area: Object Pose Estimation
Datasets: CAMERA25 / REAL275 / NOCS
Metrics: 3D IoU / 5° 2cm / 5° 5cm / 10° 2cm / 10° 5cm
Affiliations: ShanghaiTech University / Tencent / Australian National University
DVMNet: Computing Relative Pose for Unseen Objects Beyond Hypotheses
DVMNet lifts two RGB feature maps into object-centric 3D voxels with cross-view attention. Soft voxel matching and Weighted Closest Voxel weighted SVD directly recover relative rotation without enumerating poses. Mean angular error across 10 unseen CO3D categories is 19.95°; GROP Objaverse / LINEMOD errors are 20.19° / 36.82°. Inference takes around 23 ms per pair on an A100.
Chen Zhao, Tong Zhang, Zheng Dang, Mathieu Salzmann
Area: Object Pose Estimation
Datasets: CO3D / GROP-LINEMOD / GROP-Objaverse
Metrics: mean angular error / multiply-accumulate operations / pair latency
Affiliations: EPFL / ClearSpace SA
Dynamic 3D Gaussians: Tracking by Persistent Dynamic View Synthesis
Dynamic 3D Gaussians keeps Gaussian appearance and identity fixed while updating position and orientation over time. Local rigidity and isometry regularization yield dense 6-DoF trajectories from dynamic multiview rendering.
Jonathon Luiten, Georgios Kopanas, Bastian Leibe, Deva Ramanan
Area: 3D Representations / Reconstruction
Datasets: PanopticSports / Particle-NeRF
Metrics: PSNR / SSIM / LPIPS / 2D / 3D median trajectory error / tracking accuracy / survival rate
Affiliations: Carnegie Mellon University / RWTH Aachen University / Inria & Université Côte d’Azur
EDOPT: Event-Camera 6-DoF Dynamic Object Pose Tracking
EDOPT uses an incrementally updated EROS event surface as observation and 13 local pose hypotheses of a known mesh as expectations, with two asynchronous threads for event-only 6-DoF tracking. It reports offline mean errors of 0.70 cm / 2.30° and online updates above 300 Hz. No object-specific training is required, but external initialization, a known model, and continuously visible event edges are needed.
Arren J. Glover, Luna Gava, Zhichao Li, Chiara Bartolozzi
Area: Object Pose Estimation
Datasets: EDOPT synthetic dragon comparison / EDOPT five-object event pose dataset / live ATIS toy-car trial
Metrics: mean position error / mean rotation error / online update rate / tracking error over six independent-axis motions
Affiliations: Istituto Italiano di Tecnologia
Direct 3D Model-Based Object Tracking with Event Camera by Motion Interpolation
BIAM converts ROI events into absolute brightness increments and interpolates 3D model rendering across the entire motion window. Predicted images reproduce accumulation blur in real events, improving continuous tracking with perspective / fisheye cameras at speeds up to around 2 m/s.
Yufan Kang, Guillaume Caron, Ryoichi Ishikawa, Adrien Escande, Kevin Chappellet, Ryusuke Sagawa, Takeshi Oishi
Area: Object Pose Estimation
Datasets: BIAM event object-tracking dataset / YCB objects
Metrics: RMS position error / RMS rotation error / tracking duration
Affiliations: CNRS-AIST Joint Robotics Laboratory / The University of Tokyo / Inria
Real-Time 6-DoF Pose Estimation by an Event-Based Camera Using Active LED Markers
Unique LED blinking frequencies establish identities between event pixels and known marker geometry. Short-window accumulation detects LEDs and their frequencies; independent trackers update subpixel LED centers per event, and IPPE-PnP asynchronously outputs marker 6-DoF pose relative to the event camera. Desktop tracking averages 354 μs latency at 3.81 kHz. Static experiments at 2.1–4.8 m give mean errors of 34.5 mm / 0.738°, with demonstrations of approximately 10 m/s drone flight. Performance depends on active hardware, bias tuning, full LED visibility, and planar PnP geometry. Outdoor high-speed experiments do not establish preservation of static accuracy.
Gerald Ebmer, Adam Loch, Minh Nhat Vu, Roberto Mecca, Germain Haessig, Christian Hartl-Nesic, Markus Vincze, Andreas Kugi
Area: Object Pose Estimation
Datasets: custom OptiTrack active-marker accuracy test / custom indoor and outdoor quadcopter flights
Metrics: translation error / rotation error / normalized position error / tracking latency / pose output rate / static pose noise
Affiliations: TU Wien / Austrian Institute of Technology / Prophesee
FAFA: Frequency-Aware Flow-Aided Self-Supervision for Underwater Object Pose Estimation
FAFA augments PBR synthetic RGB using real-amplitude mixing / dropout, then self-supervises refinement with EMA-teacher pseudo-flow / pseudo-poses and photometric, warped-mask, feature, and point-matching consistency. ROV6D / DeepURL means are 91.39 / 73.46. It still requires known UUV CAD, coarse poses, and 1,000 real images per domain; the repository currently contains no implementation.
Jingyi Tang, Gu Wang, Zeyu Chen, Shengquan Li, Xiu Li, Xiangyang Ji
Area: Object Pose Estimation
Datasets: ROV6D / DeepURL
Metrics: ADD at 0.05d / 0.1d / ADD-S at 0.05d / 0.1d / 5° accuracy / 5cm accuracy / mean threshold accuracy
Affiliations: Tsinghua University / Peng Cheng Laboratory
Implicit Coarse-to-Fine 3D Perception for Category-Level Object Pose Estimation from Monocular RGB Image
FAP-Net learns coarse 3D awareness through geometry supervision, then refines features with pose and size constraints. Its public abstract reports gains of 14.7 and 11.4 percentage points on REAL275 10°10cm and IoU50, respectively, at 42 FPS.
Jia Li, Li Jin, Xibin Song, Yeheng Chen, Nan Li, Xueying Qin
Area: Object Pose Estimation
Datasets: CAMERA25 / REAL275 / NOCS
Metrics: 3D IoU50 / 10° 10cm / runtime
Affiliations: Samsung
FastCAD: Real-Time CAD Retrieval and Alignment from Scans and Videos
FastCAD combines object detection, upright 9-DoF alignment, and CAD embedding regression in one scene-level forward pass. Training-time embedding distillation removes the per-object encoder at test time. Scan input achieves 61.7% / 50ms, and the video pipeline 48.2% / 100ms. Reconstruction accuracy of 29.6 applies only when retrieval is restricted to the same pool as Vid2CAD.
Florian Langer, Jihong Ju, Georgi Dikov, Gerhard Reitmayr, Mohsen Ghafoorian
Area: Object Pose Estimation
Datasets: ScanNet / Scan2CAD / ShapeNet
Metrics: Scan2CAD instance alignment accuracy / class alignment accuracy / reconstruction accuracy / shape accuracy / translation error threshold / rotation error threshold / scale error threshold / inference time
Affiliations: XR Labs, Qualcomm Technologies / University of Cambridge
Fit-NGP: Fitting Object Models to Neural Graphics Primitives
Fit-NGP scans a static scene with a wrist-mounted RGB camera, trains Instant-NGP, and fits known CAD surfaces directly to the density field. Joint camera-pose optimization and normal-neighborhood sampling are key to median errors of 1.6 mm / 3.3°.
Marwan Taher, Ignacio Alzugaray, Andrew J. Davison
Area: Object Pose Estimation
Datasets: Four real tabletop scenes with 43 object instances
Metrics: median translation error / median rotation error
Affiliations: Dyson Robotics Lab, Imperial College London
Few-View Object Reconstruction with Unknown Categories and Camera Poses
FORGE uses 3D voxel features for both cross-view pose correlation and neural volume reconstruction. Rendering losses further jointly correct sparse-view camera poses at test time.
Hanwen Jiang, Zhenyu Jiang, Kristen Grauman, Yuke Zhu
Area: 3D Representations / Reconstruction
Datasets: Kubric / OmniObject3D / Google Scanned Objects
Metrics: PSNR / SSIM / LPIPS / relative rotation error / relative translation error
Affiliations: The University of Texas at Austin
FoundationPose: Unified 6D Pose Estimation and Tracking of Novel Objects
Unifies RGB-D 6D pose estimation and tracking of rigid instances unseen in training. With CAD, the model is rendered directly; without CAD, a few posed reference images fit an object-centric neural field for RGB-D rendering. Globally sampled hypotheses are refined by a Transformer RefineNet and ranked by a cross-hypothesis self-attention ScoreNet. LLMs and diffusion models only augment synthetic training data; testing uses no text input or main-network fine-tuning.
Bowen Wen, Wei Yang, Jan Kautz, Stan Birchfield
Area: Object Pose Estimation
Datasets: LINEMOD / Occluded-LINEMOD / YCB-Video / T-LESS / YCBInEOAT
Metrics: ADD AUC / ADD-S AUC / ADD-0.1d / BOP AR (VSD/MSSD/MSPD) / tracking speed
Affiliations: NVIDIA
FoundPose: Unseen Object Pose Estimation with Foundation Features
FoundPose registers DINOv2 layer-18 patch features from 800 CAD templates into 3D in under 5 minutes of onboarding per object. Online BoW retrieval selects 5 views, followed by 2D–3D matching and PnP-RANSAC. Training-free coarse / featuremetric AR is 37.2 / 42.6. The strongest 59.6-AR configuration additionally uses a MegaPose refiner trained on over 2 million synthetic images.
Evin Pınar Örnek, Yann Labbé, Bugra Tekin, Lingni Ma, Cem Keskin, Christian Forster, Tomas Hodan
Area: Object Pose Estimation
Datasets: LM-O / T-LESS / TUD-L / IC-BIN / ITODD / HB / YCB-Video
Metrics: BOP Average Recall / VSD / MSSD / MSPD / seconds per image / object onboarding time / per-object representation memory
Affiliations: Technical University of Munich / Meta Reality Labs
FreeZe: Training-free Zero-shot 6D Pose Estimation with Geometric and Vision Foundation Models
FreeZe separately normalizes and concatenates multiscale GeDi geometry and 3D-lifted DINOv2 features from CAD and RGB-D targets. Feature nearest neighbors, RANSAC, ICP, and visual symmetry selection recover pose. With CNOS masks, coarse / refined BOP AR is 64.0 / 69.3. SAR adds only 0.5 on average over ICP, and end-to-end runtime is unreported.
Andrea Caraffa, Davide Boscaini, Amir Hamza, Fabio Poiesi
Area: Object Pose Estimation
Datasets: LM-O / T-LESS / TUD-L / IC-BIN / ITODD / HB / YCB-Video
Metrics: BOP Average Recall / VSD / MSSD / MSPD
Affiliations: Fondazione Bruno Kessler / University of Trento
GenFlow: Generalizable Recurrent Flow for 6D Pose Refinement of Novel Objects
GenFlow recurrently estimates flow, certainty, pose sensitivity, and 6D pose between observations and the current CAD rendering. 3D shape-constrained correlation lookup and differentiable PnP preserve projection geometry. Under the CNOS unseen-object protocol, RGB / RGB-D mean AR is 57.8 / 67.2. The complete-design ablation reaches 70.7, 4.4 above the version without shape constraints.
Sungphill Moon, Hyeontae Son, Dongcheol Hur, Sangwook Kim
Area: Object Pose Estimation
Datasets: LM-O / T-LESS / TUD-L / IC-BIN / ITODD / HomebrewedDB / YCB-Video / MegaPose synthetic training set
Metrics: BOP Average Recall / VSD / MSSD / MSPD / render count / refinement latency
Affiliations: NAVER LABS
GeoReF: Geometric Alignment Across Shape Variation for Category-level Object Pose Refinement
GeoReF aligns observed instances and category shape priors with an HS-layer and three learned affine transforms. Cross-Cloud Transformation mixes both clouds' features and regresses rotation, translation, and size residuals. With SPD initialization, REAL275 reaches IoU75 51.8 and 5°2cm 54.4. Using only 2% of CAMERA25 training images still achieves IoU75 77.5.
Linfang Zheng, Tze Ho Elden Tse, Chen Wang, Yinghan Sun, Hua Chen, Aleš Leonardis, Wei Zhang, Hyung Jin Chang
Area: Object Pose Estimation
Datasets: NOCS-REAL275 / NOCS-CAMERA25
Metrics: 3D IoU50 / IoU75 mAP / 5°2cm / 5°5cm / 10°2cm / 10°5cm / 2cm / 5°
Affiliations: Southern University of Science and Technology / The University of Hong Kong / University of Birmingham
GigaPose: Fast and Robust Novel Object Pose Estimation via One Correspondence
GigaPose renders only 162 out-of-plane templates, predicts scale, in-plane rotation, and translation from a single 2D–2D patch correspondence, and selects pose with RANSAC. Coarse mean AR across seven BOP datasets is 26.8 at 0.4 s/image. A single detection's coarse estimate takes 48 ms, around 35× faster than MegaPose's 1.68 s; five-hypothesis GenFlow refinement reaches mean AR 60.5.
Van Nguyen Nguyen, Thibault Groueix, Mathieu Salzmann, Vincent Lepetit
Area: Object Pose Estimation
Datasets: LM-O / T-LESS / TUD-L / IC-BIN / ITODD / HomebrewedDB / YCB-Video / BOP synthetic training set
Metrics: BOP Average Recall / VSD / MSSD / MSPD / onboarding time / coarse pose latency
Affiliations: LIGM, École des Ponts / Adobe / EPFL
GS-Pose: Category-Level Object Pose Estimation via Geometric and Semantic Correspondence
GS-Pose back-projects multiview DINOv2 features onto CAD points and learns synthetic full-to-partial 3D correspondences with normalized geometry. RANSAC + Umeyama recover pose / scale of unseen real instances. With 10 CAD models per category, REAL275 results are 63.2 3D*50 / 28.8 5°5cm. The advantage is data efficiency rather than uniformly outperforming strong models trained with real supervision.
Pengyuan Wang, Takuya Ikeda, Robert Lee, Koichi Nishiwaki
Area: Object Pose Estimation
Datasets: ShapeNet / NOCS-REAL275 / Wild6D / SUN RGB-D
Metrics: 3D IoU precision / 5°5cm / 10°5cm / 15°5cm / 20°10cm / 40°20cm / 60°30cm / inference time
Affiliations: Technical University of Munich / Woven by Toyota
HiPose: Hierarchical Binary Surface Encoding and Correspondence Pruning for RGB-D 6DoF Object Pose Estimation
HiPose predicts hierarchical binary surface codes for each RGB-D point. Coarse point-to-surface correspondences are narrowed bit by bit, rejecting outliers according to the current pose before final Kabsch fitting. LM-O ADD(-S) recall is 89.6 and the three-dataset BOP mean score is 84.6. Pose inference takes 0.075 s, or 0.16 s with detection, without rendering-based refinement.
Yongliang Lin, Yongzhi Su, Praveen Nathan, Sandeep Inuganti, Yan Di, Martin Sundermeyer, Fabian Manhardt, Didier Stricker, Jason Rambach, Yu Zhang
Area: Object Pose Estimation
Datasets: LM-O / YCB-Video / T-LESS / BOP PBR
Metrics: ADD(-S) recall / AUC of ADD-S / AUC of ADD(-S) / BOP score / runtime
Affiliations: Zhejiang University / German Research Center for Artificial Intelligence (DFKI) / RPTU Kaiserslautern-Landau / Technical University of Munich / Google
HOC-Search: Efficient CAD Model and Pose Retrieval from RGB-D Scans
HOC-Search organizes categories, CAD models, and 9-DoF poses into a hierarchical search tree. MCTS approximates expensive exhaustive render-and-compare search, while joint pose refinement focuses on promising candidates.
Stefan Ainetter, Sinisa Stekovic, Friedrich Fraundorfer, Vincent Lepetit
Area: Object Pose Estimation
Datasets: ScanNet / Scan2CAD
Metrics: Top-1 / Top-5 CAD retrieval accuracy / Chamfer Distance / speedup factor
Affiliations: Graz University of Technology / LIGM, École des Ponts, Université Gustave Eiffel, CNRS
HouseCat6D - A Large-Scale Multi-Modal Category Level 6D Object Perception Dataset with Household Objects in Realistic Scenarios
HouseCat6D provides 194 objects, 10 categories, 41 real scenes, 23.5k frames, around 160k poses, and roughly 10M grasp annotations. External infrared tracking, calibration, and bundle adjustment control point-level annotation errors to 1.35–1.73 mm. VI-Net's overall 3D25/50 is 80.7/56.4, with strong transparent-glass results but continued difficulty on boxes / remotes.
HyunJun Jung, Shun-Cheng Wu, Patrick Ruhkamp, Guangyao Zhai, Hannah Schieber, Giulia Rizzoli, Pengyuan Wang, Hongcheng Zhao, Lorenzo Garattoni, Sven Meier, Daniel Roth, Nassir Navab, Benjamin Busam
Area: Object Pose Estimation
Datasets: HouseCat6D
Metrics: 3D IoU at 25% / 50% / point-wise annotation RMSE / grasp coverage rate / object success rate / real-world grasp success rate
Affiliations: Technical University of Munich / Munich Center for Machine Learning / FAU Erlangen-Nürnberg / University of Padova / Toyota Motor Europe / 3dwe.ai
IFFNeRF: Initialisation Free and Fast 6DoF Pose Estimation from a Single Image and a NeRF Model
IFFNeRF samples surfaces from a NeRF density field and generates parameterized rays. DINOv2 image features and attention retrieve rays pointing toward the true camera center; weighted linear least squares recovers pose without initialization in around 29 ms.
Matteo Bortolon, Theodore Tsesmelis, Stuart James, Fabio Poiesi, Alessio Del Bue
Area: Object Pose Estimation
Datasets: NeRF Synthetic / Tanks and Temples
Metrics: mean angular error / mean translation error / runtime / memory
Affiliations: Istituto Italiano di Tecnologia / Fondazione Bruno Kessler / University of Trento / Durham University
Towards Co-Evaluation of Cameras, HDR, and Algorithms for Industrial-Grade 6DoF Pose Estimation
IPD provides 20 industrial parts, 2,300 physical scenes, 13 cameras, and over 100,000 object views. Robot Consistency evaluates MVD without millimeter-accurate manual ground truth. Synthetic validation gives an error bias of 0.13 mm versus 0.47 mm with 1 mm annotation jitter. Basler-HR is the most stable; Mertens HDR under spotlights achieves 3.061 mm MVD / 0.77 recall.
Agastya Kalra, Guy Stoppi, Dmitrii Marin, Vage Taamazyan, Aarrushi Shandilya, Rishav Agarwal, Anton Boykov, Tze Hao Chong, Michael Stark
Area: Object Pose Estimation
Datasets: Industrial Plenoptic Dataset (IPD)
Metrics: Maximum Vertex Distance / robot consistency error / precision / recall
Affiliations: Intrinsic Innovation LLC
Joint-Loss Enhanced Self-Supervised Learning for Refinement-Coupled Object 6D Pose Estimation
An Iterative Annotation Resolver builds supervision from RGB-D data without 6D annotations. Pixelwise weighted dense fusion initializes pose, followed by refinement trained with a differentiable renderer and joint loss. The authors' abstract claims accuracy close to supervised methods, without specifying datasets or numerical results.
Fengjun Mu, Shixiang Sun, Rui Huang, Chaobin Zou, Wenjiang Li, Huayi Zhan, Hong Cheng
Area: Object Pose Estimation
Datasets: Unannotated RGB-D object observations (dataset not disclosed in public abstract)
Metrics: 6D pose accuracy (public abstract gives qualitative comparison only)
Affiliations: University of Electronic Science and Technology of China / Changhong
Learning Better Keypoints for Multi-Object 6DoF Pose Estimation
KeyGNet replaces fixed FPS / BBox keypoints with offline target design. EdgeConv learns shared 3D keypoints from visible colored object surfaces within a dataset. Wasserstein loss aligns voting distributions across keypoints, and dispersion loss prevents geometric collapse. Replacing keypoints in PVNet, PVN3D, and RCVPose raises MIMO mean AR across seven BOP core datasets by 11.3, 12.0, and 10.0 points; PVN3D LMO ADD(S) rises from 53.9 to 70.3. Benefits center on closing the SISO→MIMO gap, but ground-truth poses, intrinsics, known geometry, visible surfaces, and separate training per dataset / voting representation are required.
Yangzheng Wu, Michael Greenspan
Area: Object Pose Estimation
Datasets: LINEMOD-Occlusion / YCB-Video / T-LESS / TUD-L / IC-BIN / ITODD / HomebrewedDB
Metrics: BOP Average Recall / VSD recall / MSSD recall / MSPD recall / ADD(S) / ADD(S) AUC / SISO-to-MIMO performance gap / training time
Affiliations: Queen's University
L4D-Track: Language-to-4D Modeling Towards 6-DoF Tracking and Shape Reconstruction in 3D Point Cloud Stream
L4D-Track combines pairwise implicit 3D pose / shape fields with multilayer language alignment to causally track 6-DoF motion and reconstruct objects from point cloud streams. NOCS-REAL275 IoU75 / 5°5cm is 76.0 / 56.2; cross-domain YCB-Video without fine-tuning yields mean ADD / ADD-S of 80.4 / 86.1. NOCS / YCB speeds are 20.45 / 19.28 FPS, but the tracking protocol requires first-frame ground-truth pose and size.
Jingtao Sun, Yaonan Wang, Mingtao Feng, Yulan Guo, Ajmal Mian, Mike Zheng Shou
Area: Object Pose Estimation
Datasets: NOCS-REAL275 / YCB-Video
Metrics: 3D IoU25 / IoU50 / IoU75 / 5°2cm / 5°5cm / 10°2cm / 10°5cm / rotation / translation error / ADD / ADD-S / Chamfer Distance / FPS
Affiliations: Hunan University / National University of Singapore / Xidian University / Sun Yat-sen University / The University of Western Australia
LaPose: Laplacian Mixture Shape Modeling for RGB-Based Category-Level Object Pose Estimation
LaPose predicts Laplace means / variances for NOCS through DINOv2 / ConvNet branches, allowing learned PnP to use local shape uncertainty. Translation / size are normalized by the ground-truth 3D box diagonal, with absolute scale separately estimated by MobileNet. A single model achieves REAL275 70.7 NIoU25 and 57.4 normalized 10°0.5d, but absolute IoU75 is only 2.6.
Ruida Zhang, Ziqin Huang, Gu Wang, Chenyangguang Zhang, Yan Di, Xingxing Zuo, Jiwen Tang, Xiangyang Ji
Area: Object Pose Estimation
Datasets: NOCS-CAMERA25 / NOCS-REAL275
Metrics: Normalized 3D IoU mAP / normalized 10°0.2d / 10°0.5d / absolute 3D IoU mAP / 10°10cm / rotation threshold accuracy / normalized translation threshold accuracy
Affiliations: Tsinghua University / Technical University of Munich / California Institute of Technology
LFM-3D: Learnable Feature Matching Across Wide Baselines Using 3D Signals
LFM-3D encodes NOCS or monocular depth as 3D positional signals for keypoints, injecting them into a graph matching network to improve wide-baseline correspondences and relative rotation estimation between object images.
Arjun Karpur, Guilherme Mendeleh Perrotta, Ricardo Martin-Brualla, Howard Zhou, André Araujo
Area: Feature Matching / Correspondences
Datasets: Google Scanned Objects / Objectron
Metrics: matching precision / recall / relative rotation Acc@5° / 10° / 15°
Affiliations: Google Research / University of Campinas
Living Scenes: Multi-object Relocalization and Reconstruction in Changing 3D Environments
MoRE² uses a shared SE(3)-equivariant encoder–DeepSDF decoder for cross-scan instance matching, 6-DoF registration, and object reconstruction, jointly optimizing the pose graph and latent geometry. FlyingShape / 3RScan MR is 74.39 / 49.07 and MRR is 62.00 / 40.74. Advantages persist with noisy Mask3D segmentations, but instance segmentation is required, and nonrigidity or ambiguity between identical instances is not addressed.
Liyuan Zhu, Shengyu Huang, Konrad Schindler, Iro Armeni
Area: Object Pose Estimation
Datasets: FlyingShape / 3RScan / ShapeNet
Metrics: MR Recall / MRR Recall / instance matching recall / registration recall / median rotation error / RMSE / Chamfer distance / volumetric IoU / SDF recall
Affiliations: Stanford University / ETH Zurich
LocPoseNet: Robust Location Prior for Unseen Object Pose Estimation
LocPoseNet combines bidirectional multiscale reference kernels with decoupled scale-robust and scale-aware heads to predict unseen-object centers and scales, providing robust location priors for translation and subsequent rotation refinement.
Chen Zhao, Yinlin Hu, Mathieu Salzmann
Area: Object Pose Estimation
Datasets: LINEMOD / GenMOP / Synthetic LINEMOD robustness set
Metrics: mAP@[.5:.95] / ADD-0.1d
Affiliations: EPFL CVLab / ClearSpace SA / Magic Leap
Effects of Markers in Training Datasets on the Accuracy of 6D Pose Estimation
Paired synthetic scenes isolate marker shortcuts in training images, comparing fixed markers, two-stage LaMa inpainting, fully random markers, and mixed fixed / random augmentation. On markerless Teabox tests, fixed-marker training yields ADD-(S) 76.6, 8.2 points below the markerless baseline's 84.8; inpainting restores 84.8. Five-object YCB-V and ZebraPose checks broadly support limited pose degradation after inpainting. Real averages exclude severely failed mug / wood-block cases, and multiple-seed intervals are absent. Label-noise sweeps show that translation errors above 4 mm or rotation errors around 2.5° quickly erode precise-ground-truth benefits.
Janis Rosskamp, Rene Weller, Gabriel Zachmann
Area: Object Pose Estimation
Datasets: synthetic teabox paired-marker dataset / real teabox marker / markerless dataset / YCB-Video five-object subset
Metrics: AUC of ADD-S / AUC of ADD-(S) / ADD-(S) at 10% object diameter / translation label error / rotation label error
Affiliations: University of Bremen
MatchU: Matching Unseen Objects for 6D Pose Estimation from RGB-D Images
MatchU fuses CAD geometry, observed depth, and RGB texture into color-aware rotation-invariant 3D descriptors, then uses matching, RANSAC, and optional ICP for 6D poses of objects unseen in training. Single-hypothesis BOP-5 mean AR without refinement is 50.5 at 0.07 s; multiple hypotheses reach 57.8, and refinement reaches 64.7. RGB disambiguation improves correspondences, while multiple hypotheses improve global robustness.
Junwen Huang, Hao Yu, Kuan-Ting Yu, Nassir Navab, Slobodan Ilic, Benjamin Busam
Area: Object Pose Estimation
Datasets: Google Scanned Objects / LM-O / T-LESS / TUD-L / IC-BIN / YCB-Video
Metrics: BOP Average Recall / VSD / MSSD / MSPD / ADD-0.1d / runtime
Affiliations: Technical University of Munich / Munich Center for Machine Learning / XYZ Robotics / 3dwe.ai
MFOS: Model-Free & One-Shot Object Pose Estimation
For CAD-free one-shot RGB 6D pose estimation of unseen rigid instances, a posed reference image and coarse 3D cuboid supply the priors. A CroCo-initialized ViT fuses the reference image with a proxy-shape pointmap to predict dense query-to-unseen-cuboid 2D–3D correspondences and confidence, followed by RANSAC-PnP.
JongMin Lee, Yohann Cabon, Romain Brégier, Sungjoo Yoo, Jerome Revaud
Area: Object Pose Estimation
Datasets: LINEMOD / OnePose / OnePose-LowTexture / Amazon Berkeley Objects (ABO)
Metrics: ADD(-S)-0.1d / Proj2D (5px) / 1cm-1deg / 3cm-3deg / 5cm-5deg / inference time
Affiliations: Seoul National University / Naver Labs Europe
Learning Instance-Aware Correspondences for Robust Multi-Instance Point Cloud Registration in Cluttered Scenes
MIRETR iteratively learns instance-aware superpoint features and masks, expanding coarse matches into per-instance dense correspondences and SE(3) hypotheses without expensive multimodel fitting. The full model's mean F1 on Scan2CAD / ROBI / ShapeNet is 93.40 / 39.80 / 94.44. On ROBI, removing both instance-aware designs lowers F1 to 18.65, and reverting to global attention gives only 10.96.
Zhiyuan Yu, Zheng Qin, Lintao Zheng, Kai Xu
Area: Object Pose Estimation
Datasets: Scan2CAD / ROBI / ShapeNet
Metrics: Inlier Ratio / Mean Recall / Mean Precision / Mean F1 / instance mask mIoU
Affiliations: National University of Defense Technology / Defense Innovation Institute, Academy of Military Sciences
MOTPose: Multi-Object 6D Pose Estimation for Dynamic Video Sequences Using Attention-Based Temporal Fusion
Within a DETR-style multi-object network, MOTPose's TEFM fuses historical object embeddings and TOFM fuses keypoints with pose parameters. Compared with its single-frame version, it reduces missed detections and pose outliers on dynamic SynPick while retaining competitive accuracy on YCB-Video.
Arul Selvam Periyasamy, Sven Behnke
Area: Object Pose Estimation
Datasets: SynPick / SynPick-Ext / YCB-Video
Metrics: ADD AUC / ADD-S AUC / ADD(-S) AUC / cardinality error / false-negative rate
Affiliations: University of Bonn
MRC-Net: 6-DoF Pose Estimation with MultiScale Residual Correlation
MRC-Net quantizes rotations into 4,608 prototypes for coarse pose prediction, then renders CAD and regresses residuals from three-scale real–render correlations. Symmetry-aware soft labels handle multiple solutions. PBR-only training yields mean AR 63.3 on four BOP datasets. The T-LESS ablation improves from parallel MultiTask's 64.9 to 77.1, with around 61 ms per object.
Yuelong Li, Yafei Mao, Raja Bala, Sunil Hadap
Area: Object Pose Estimation
Datasets: T-LESS / ITODD / YCB-Video / LM-O
Metrics: BOP Average Recall / VSD / MSSD / MSPD / ADD(-S) / AUC of ADD-S / AUC of ADD(-S) / single-object latency
Affiliations: Amazon Inc.
MV-ROPE: Multi-View Constraints for Robust Category-Level Object Pose and Size Estimation
MV-ROPE recovers camera motion and metric depth from video with scale-aware dense monocular SLAM. NOCS–depth registration estimates keyframe object poses, while IoU association and an object-level pose graph fuse multiview constraints.
Jiaqi Yang, Yucong Chen, Xiangting Meng, Chenxin Yan, Min Li, Ran Cheng, Lige Liu, Tao Sun, Laurent Kneip
Area: Object Pose Estimation
Datasets: NOCS REAL / MEREAL
Metrics: 3D IoU mAP / 5°2cm mAP / 5°5cm mAP / mean rotation error / mean translation error
Affiliations: ShanghaiTech University / Midea RoboZone
Perception Through Cognitive Emulation: A Second Iteration of NaivPhys4RP for Learningless and Safe Recognition and 6D-Pose Estimation of (Transparent) Objects
NaivPhys4RP generates possible worlds from context, ontology, and naive physics, predicts their evolution in a simulator, and filters belief particles with limited observations to explain transparent-object identity and pose in a white-box system. It is a system proof of concept without a quantitative BOP ranking.
Franklin Kenghagho Kenfack, Michael Neumann, Patrick Mania, Michael Beetz
Area: World Models
Datasets: Transparent-object cognitive emulation demonstrations
Metrics: qualitative recognition and 6D pose demonstration
Affiliations: University of Bremen
NCRF: Neural Contact Radiance Fields for Free-Viewpoint Rendering of Hand-Object Interaction
NCRF jointly models contact optimization fields, canonical hand–object motion fields, and radiance fields, allowing contact and multiview rendering to correct hand and object poses under occlusion.
Zhongqun Zhang, Jifei Song, Eduardo Pérez-Pellitero, Yiren Zhou, Hyung Jin Chang, Aleš Leonardis
Area: 3D Representations / Reconstruction
Datasets: HO3D / DexYCB
Metrics: PSNR / SSIM / LPIPS / hand MPJPE / ADD-0.1D / intersection volume
Affiliations: University of Birmingham / Huawei Noah’s Ark Lab
Marrying NeRF with Feature Matching for One-Step Pose Estimation
The method renders NeRF RGB / depth from an initial pose, obtains LoFTR 2D–2D matches, lifts them into 2D–3D correspondences, and performs a single RANSAC-PnP correction. 3D consistency mining and keypoint-guided refinement address spurious geometry and occlusion, respectively.
Ronghan Chen, Yang Cong, Yu Ren
Area: Object Pose Estimation
Datasets: NeRF Synthetic / LLFF
Metrics: rotation correctness / translation correctness / mean rotation error / mean translation error / FPS
Affiliations: Shenyang Institute of Automation, Chinese Academy of Sciences / South China University of Technology
NeRF-Feat: 6D Object Pose Estimation using Feature Rendering
NeRF-Feat uses a NeRF density field to establish a CAD-free instance reference frame. Bidirectional contrastive learning aligns 3D surface features with query CNN features, and 2D–3D correspondences with PnP recover pose.
Shishir Reddy Vutukur, Heike Brock, Benjamin Busam, Tolga Birdal, Andreas Hutter, Slobodan Ilic
Area: Object Pose Estimation
Datasets: LINEMOD / LM-Occlusion / T-LESS
Metrics: ADD / ADD-S / BOP Average Recall
Affiliations: Technical University of Munich / Siemens AG / Imperial College London
NeuSurfEmb: A Complete Pipeline for Dense Correspondence-Based 6D Object Pose Estimation without CAD Models
NeuSurfEmb builds renderable object surfaces from a few real RGB images using SfM, general segmentation, and NeuS2. Automatically generated cut-and-paste training images train a SurfEmb dense-correspondence network. Exact CAD and a manually built PBR pipeline are unnecessary, but per-object reconstruction and training remain required.
Francesco Milano, Jen Jen Chung, Hermann Blum, Roland Siegwart, Lionel Ott
Area: Object Pose Estimation
Datasets: LINEMOD-Occlusion / NeuSurfEmb five-object real-world benchmark
Metrics: BOP Average Recall / ADD(-S) recall / 5 cm, 5° recall / forward Chamfer distance
Affiliations: ETH Zurich / The University of Queensland
NOPE: Novel Object Pose Estimation from a Single Image
Rather than synthesizing pixels, NOPE predicts discriminative embeddings for 342 relative views from one reference and retrieves against the query embedding, recovering relative rotation and multimodal ambiguity. On unseen ShapeNet categories, top-1 Acc30 / median error is 59.8 / 24.4°, and top-5 is 87.1 / 6.6°. Single-reference T-LESS VSD recall is 46.49, with 40.5 Acc30 under 25% occlusion.
Van Nguyen Nguyen, Thibault Groueix, Georgy Ponimatkin, Yinlin Hu, Renaud Marlet, Mathieu Salzmann, Vincent Lepetit
Area: Object Pose Estimation
Datasets: ShapeNet / T-LESS
Metrics: Acc30 / median angular error / VSD recall / occlusion robustness / memory / runtime
Affiliations: LIGM, École des Ponts / Adobe / Magic Leap / Valeo.ai / EPFL
Omni6D: Large-Vocabulary 3D Object Dataset for Category-Level 6D Object Pose Estimation
Omni6D synthesizes over 800,000 RGB-D observations in Replica scenes from 4,688 scanned textured meshes across 166 categories, with instance-level symmetry along three axes and unseen-category tests. Baselines use ground-truth masks. Translation and size generalize, but rotation degrades sharply under large-vocabulary and cross-category settings.
Mengchen Zhang, Tong Wu, Tai Wang, Tengfei Wang, Ziwei Liu, Dahua Lin
Area: Object Pose Estimation
Datasets: Omni6D / Omni6Dout / Omni6D-Real / REAL275 / CAMERA25
Metrics: 3D IoU at 50% / 75% / 5°2cm / 5°5cm / 10°2cm / 10°5cm / rotation threshold accuracy / translation threshold accuracy / symmetry-aware pose error
Affiliations: Zhejiang University / Shanghai Artificial Intelligence Laboratory / The Chinese University of Hong Kong / Nanyang Technological University
Omni6DPose: A Benchmark and Model for Universal 6D Object Pose Estimation and Tracking
Omni6DPose comprises real ROPE data across 149 categories (332K images, 1.5M+ annotations, 581 instances), synthetic SOPE (475K images, 5M+ annotations, 4,162 instances), and aligned PAM meshes. GenPose++ combines DINOv2 / PointNet++ conditional diffusion with clustering aggregation to reach 15.1 VUS@5°5cm. Main-table results use ground-truth segmentation, and absolute accuracy shows that large-vocabulary pose remains difficult.
Jiyao Zhang, Weiyao Huang, Bo Peng, Mingdong Wu, Fei Hu, Zijian Chen, Bo Zhao, Hao Dong
Area: Object Pose Estimation
Datasets: ROPE / SOPE / PAM
Metrics: AUC at 3D IoU thresholds / Volume Under Surface at rotation / translation thresholds / tracking mIoU / rotation error / translation error / FPS
Affiliations: Peking University / Waseda University / Shanghai Jiao Tong University / Beijing Academy of Artificial Intelligence
OmniNOCS: A Unified NOCS Dataset and Model for 3D Lifting of 2D Objects
OmniNOCS unifies 10 sources into canonical NOCS data covering 97 categories, 380k images, and over 450k instances. NOCSformer predicts masks, NOCS, metric size, rotation, and centroids from external 2D boxes. On held-out NOCS-Real275, 3D IoU mAP@25 / 50 is 43.5 / 10.6, exceeding transferred Cube R-CNN but far below the target-domain-supervised NOCS model's 79.6 / 72.4.
Akshay Krishnan, Abhijit Kundu, Kevis-Kokitsi Maninis, James Hays, Matthew Brown
Area: Object Pose Estimation
Datasets: OmniNOCS / KITTI / SUN RGB-D / Objectron / nuScenes / Hypersim / ARKitScenes / Cityscapes 3D / Virtual KITTI / NOCS-Real275 / Waymo Open Dataset
Metrics: NOCS mAE / NOCS mPSNR / mask mIoU / 3D IoU mAP / mATE / mAOE / mASE / orientation threshold accuracy
Affiliations: Google Research / Google DeepMind / Georgia Institute of Technology
OP-Align: Object-level and Part-level Alignment for Self-supervised Category-level Articulated Object Pose Estimation
OP-Align uses E2PN's 60 rotation anchors to align a single partial point cloud globally to a category-shared canonical reconstruction. Predicted joint directions, pivots, and states align parts separately, while distance differences self-supervise part segmentation. It reaches 41 FPS and outperforms EAP on synthetic data, but real-data joint strict mAP is only 12.57. It requires known part counts, joint types, and training ground truth to calibrate evaluation frames.
Yuchen Che, Ryo Furukawa, Asako Kanezaki
Area: Object Pose Estimation
Datasets: EAP synthetic articulated-object dataset / OP-Align real-world articulated-object dataset / HOI4D / Shape2Motion
Metrics: part segmentation IoU / part rotation / translation error / joint direction / pivot error / multi-threshold mAP / FPS / GPU memory
Affiliations: Tokyo Institute of Technology / Accenture Japan Ltd.
Open-Vocabulary Object 6D Pose Estimation
Oryon uses text prompts to jointly locate targets, learn local matches, and recover relative SE(3) from anchor / query RGB-D images of different scenes, without CAD or object video. With predicted masks, REAL275 / Toyota-Light BOP AR is 32.2 / 30.3; oracle masks give 46.5 / 34.1, showing that open-vocabulary segmentation remains a major performance limit.
Jaime Corsetti, Davide Boscaini, Changjae Oh, Andrea Cavallaro, Fabio Poiesi
Area: Object Pose Estimation
Datasets: ShapeNet6D / REAL275 / Toyota-Light
Metrics: BOP Average Recall / VSD / MSSD / MSPD / ADD(S)-0.1d / mask mIoU
Affiliations: Fondazione Bruno Kessler / University of Trento / Queen Mary University of London / Idiap Research Institute / EPFL
Robust Category-Level 3D Pose Estimation From Diffusion-Enhanced Synthetic Data
P3D-Diffusion fixes category-level 3D rotation labels with a renderer and augments appearance using Canny-edge-conditioned diffusion, DTD texture priors, and category-independent random backgrounds. CC3D adapts synthetic features to unlabeled real images through neural-mesh vertex correspondences. On PASCAL3D+, 10% real rotation labels yield 30° / 10° accuracy of 86.7 / 62.4 and MedErr 8.4°; 50% labels yield 90.7 / 71.4 / 6.9°. Translation / scale are not estimated, and diffusion-induced pose-label drift is not directly quantified.
Jiahao Yang, Wufei Ma, Angtian Wang, Xiaoding Yuan, Alan Yuille, Adam Kortylewski
Area: Object Pose Estimation
Datasets: PASCAL3D+ / OOD-CV / P3D-Diffusion synthetic dataset / Describable Texture Dataset
Metrics: rotation accuracy below 30 degrees / rotation accuracy below 10 degrees / median angular error / OOD nuisance-wise rotation accuracy
Affiliations: Peking University / Johns Hopkins University / University of Freiburg / Max Planck Institute for Informatics
PACE: A Large-Scale Dataset with Pose Annotations in Cluttered Environments
PACE provides 300 real cluttered videos, 55K frames, and 258K pose annotations spanning 238 objects, 43 categories, and rigid / articulated objects, plus 100K PACE-Sim frames. PPF reaches 52.2 BOP AR on known instances versus only 4.4–9.0 for learned methods. Controlled experiments separate performance collapses caused by sim-to-real transfer and scaling to the full object set.
Yang You, Kai Xiong, Zhening Yang, Zhengxiang Huang, Junwei Zhou, Ruoxi Shi, Zhou Fang, Adam W. Harley, Leonidas Guibas, Cewu Lu
Area: Object Pose Estimation
Datasets: PACE / PACE-Sim / NOCS-REAL275
Metrics: BOP Average Recall / VSD / MSSD / MSPD / 3D IoU AP / angle-threshold AP / translation-threshold AP / ADD AUC / ADD-S AUC / 5°5cm tracking accuracy
Affiliations: Stanford University / Shanghai Jiao Tong University / Horizon Robotics / UC San Diego
PEGASUS: Physically Enhanced Gaussian Splatting Simulation System for 6DoF Object Pose Dataset Generation
PEGASUS represents objects and environments with 3D Gaussians for appearance and low-poly meshes for collision. PyBullet generates natural rigid-body trajectories, which are rendered and exported as BOP-format RGB, depth, masks, boxes, and 6-DoF poses.
Lukas Meyer, Floris Erich, Yusuke Yoshiyasu, Marc Stamminger, Noriaki Ando, Yukiyasu Domae
Area: 3D Representations / Reconstruction
Datasets: Ramen / PEGASet / YCB-V assets / PEGASUS-generated BOP datasets
Metrics: dataset generation time / robotic picking success
Affiliations: Friedrich-Alexander-Universität Erlangen-Nürnberg / National Institute of Advanced Industrial Science and Technology
Physics-Based Rigid Body Object Tracking and Friction Filtering From RGB-D Videos
The method uses differentiable rigid-body simulation as the state transition of an Extended Kalman Filter, jointly tracking position, SO(3) rotation, velocity, and friction coefficients between RGB-D pose measurements.
Rama Krishna Prasad Kandukuri, Michael Strecke, Joerg Stueckler
Area: World Models
Datasets: Synthetic sliding and collision sequences / Real RGB-D sliding sequences / YCB object models
Metrics: rotation error / recall / position error / recall / linear / angular velocity error / friction coefficient error
Affiliations: Embodied Vision Group, Max Planck Institute for Intelligent Systems
Shape-guided Configuration-aware Learning for Endoscopic-image-based Pose Estimation of Flexible Robotic Instruments
PoseFlex uses attention from per-part 3D shape priors to guide endoscopic RGB features. An initial four-angle configuration creates dynamic priors for refinement, while a matrix-Fisher distribution outputs angles and uncertainty. On 12,473 private ex-vivo pairs, G3 α/β/γ/δ mean errors are 7.08°/4.66°/1.62°/11.68°. It does not predict standard 6-DoF translation.
Yiyao Ma, Kai Chen, Hon-Sing Tong, Ruofeng Wei, Yui-Lun Ng, Ka-Wai Kwok, Qi Dou
Area: Object Pose Estimation
Datasets: PoseFlex ex-vivo flexible-robot dataset
Metrics: mean angular error / median angular error / accuracy below 5° / accuracy below 10° / matrix-Fisher uncertainty
Affiliations: The Chinese University of Hong Kong / Agilis Robotics Limited / The University of Hong Kong
PS6D: Point Cloud Based Symmetry-Aware 6D Object Pose Estimation in Robot Bin-Picking
PS6D predicts poses pointwise with attention-based point cloud features, a symmetry-aware rotation loss, and a center-distance-weighted translation loss. Two-stage clustering first separates centers, then rotations, handling intersecting elongated parts and multiple symmetry modes. The paper also reports 91.7% real bin-picking success for one part category.
Yifan Yang, Zhihao Cui, Qianyi Zhang, Jingtai Liu
Area: Object Pose Estimation
Datasets: Siléane / IPA Bin-Picking Dataset / PS6D industrial dataset / Real Mech-Eye / Fanuc bin-picking trials
Metrics: instance F1 / Recall / bin-picking success rate
Affiliations: Nankai University / Mech-Mind
RelPose++: Recovering 6D Poses from Sparse-view Observations
RelPose++ updates pairwise rotation energy distributions with a multiview Transformer and regresses translation in look-at-centered coordinates, recovering a consistent set of camera 6D poses from sparse object views.
Amy Lin, Jason Y. Zhang, Deva Ramanan, Shubham Tulsiani
Area: Visual Localization / SLAM
Datasets: CO3D v2 / Objectron
Metrics: joint rotation Acc@15° / camera center Acc@0.2 / translation Acc@0.1
Affiliations: Carnegie Mellon University
RGBManip: Monocular Image-Based Robotic Manipulation Through Active Object Pose Estimation
RGBManip schedules active observation and manipulation with a high-level RL policy. Kinematics-guided multiview depth-aware features estimate category-level poses, while impedance control absorbs residual errors. Its main results are success rates on six robotic tasks rather than a standalone pose benchmark.
Boshi An, Yiran Geng, Kai Chen, Xiaoqi Li, Qi Dou, Hao Dong
Area: Robot Perception & Manipulation
Datasets: PartNet-Mobility / ShapeNet / SAPIEN manipulation benchmark
Metrics: manipulation success rate / pose error versus number of views
Affiliations: Peking University / Beijing Academy of Artificial Intelligence / The Chinese University of Hong Kong
Pseudo-keypoint RKHS Learning for Self-supervised 6DoF Pose Estimation
RKHSPose first learns radial keypoint regression / CNN voting on PBR synthetic RGB-D. Pseudo-poses from real images generate synthetic-over-real pairs, and dense MMD with a learnable RKHS kernel aligns multilayer features. Using 320 unlabeled real images improves LM-O ADD(S) from 55.7 to 68.7. Known CAD, a closed object set, and adaptation to each target domain remain necessary.
Yangzheng Wu, Michael Greenspan
Area: Object Pose Estimation
Datasets: LINEMOD / Occlusion LINEMOD / YCB-Video / T-LESS / TUD-L / ITODD / HomebrewedDB
Metrics: ADD(S) / ADD(S) AUC / ADD-S AUC / BOP Average Recall / VSD / MSSD / MSPD / inference time
Affiliations: Queen’s University
SAM-6D: Segment Anything Model Meets Zero-Shot 6D Object Pose Estimation
For zero-shot RGB-D 6D pose estimation of instances unseen in training, target CAD renderings supply the object prior. SAM/FastSAM proposes masks, filtered by DINOv2 semantic, appearance, and geometric scores. Coarse-to-fine 3D–3D matching with background tokens and weighted SVD estimates all target instance poses.
Jiehong Lin, Lihua Liu, Dekun Lu, Kui Jia
Area: Object Pose Estimation
Datasets: BOP core datasets / LM-O / T-LESS / TUD-L / IC-BIN / ITODD / HB / YCB-V
Metrics: segmentation mAP (IoU 0.50-0.95) / BOP AR (VSD/MSSD/MSPD) / ADD(-S)
Affiliations: DexForce Co., Ltd. / The Chinese University of Hong Kong, Shenzhen / South China University of Technology
Scalable Network and Adaptive Refinement Module for 6D Pose Estimation of Diverse Industrial Components
EfficientNet and a shallow bidirectional FPN estimate initial poses from RGB; observed depth supports adaptive point-cloud registration refinement. Network scale and refinement budgets share a scaling parameter, with convergence-based stopping.
Kun Qian, Mustafa Suphi Erden, Xianwen Kong
Area: Object Pose Estimation
Datasets: LINEMOD / Solar cell industrial dataset / Primary lens industrial dataset
Metrics: ADD(-S) / 5°1cm accuracy / FPS / refinement iterations
Affiliations: University of Manchester / Heriot-Watt University
SD-Net: Symmetric-Aware Keypoint Prediction and Domain Adaptation for 6D Pose Estimation in Bin-Picking Scenarios
SD-Net designs equivalent keypoint sets for each object symmetry type. Pointwise regression, instance clustering, keypoint filtering, and Hough voting recover piled-part poses; teacher–student learning and Semi-Chamfer geometric filtering adapt the synthetic model to unlabeled real point clouds.
Ding-Tao Huang, En-Te Lin, Lipeng Chen, Li-Fu Liu, Long Zeng
Area: Object Pose Estimation
Datasets: Siléane / Parametric dataset / Real Fanuc bin-picking trials
Metrics: symmetry-aware Average Precision / robotic grasping success
Affiliations: Tsinghua Shenzhen International Graduate School / Tencent Robotics X
Confronting Ambiguity in 6D Object Pose Estimation via Score-Based Diffusion on SE(3)
The method learns p(R,t|I) directly, defining score diffusion on SE(3) rather than independently diffusing rotation and translation. A surrogate Stein score avoids expensive Jacobians. On T-LESS, the SE(3) model achieves MSPD / R@5 of 93.16 / 86.94. Sampling runs at 20 FPS with 100 steps and 250 FPS with 5 steps.
Tsu-Ching Hsiao, Hao-Wei Chen, Hsuan-Kung Yang, Chun-Yi Lee
Area: Object Pose Estimation
Datasets: SYMSOL / SYMSOL-T / T-LESS
Metrics: distribution spread / MSPD / MSSD / VSD / R@2 / R@5 / R@10 / T@2 / T@5 / T@10 / FPS
Affiliations: Elsa Lab, National Tsing Hua University
SecondPose: SE(3)-Consistent Dual-Stream Feature Fusion for Category-Level Pose Estimation
SecondPose fuses DINOv2 category semantics with hierarchical PPF instance geometry pointwise into an approximately SE(3)-consistent representation, then separately regresses rotation, translation, and size. REAL275 IoU75 / 5°2cm is 49.7 / 56.2, and HouseCat6D IoU25/50 is 83.7/66.1. Removing semantics lowers strict 5°2cm to 51.1.
Yamei Chen, Yan Di, Guangyao Zhai, Fabian Manhardt, Chenyangguang Zhang, Ruida Zhang, Federico Tombari, Nassir Navab, Benjamin Busam
Area: Object Pose Estimation
Datasets: NOCS-REAL275 / NOCS-CAMERA25 / HouseCat6D
Metrics: 3D IoU75 mAP / 5°2cm / 5°5cm / 10°2cm / 10°5cm / HouseCat6D IoU25 / IoU50 / FPS
Affiliations: Technical University of Munich / Munich Center for Machine Learning / Google / Tsinghua University / 3dwe.ai
Sim-to-Real Object Pose Estimation for Random Bin Picking
Target CAD and a physics simulator automatically generate random piles, instance segmentations, and pose annotations. Training uses only synthetic point clouds, followed by deployment to real bins without additional domain adaptation and integration into a robot-picking loop. The public abstract reports competitive synthetic recall / accuracy and improved real grasp success without numerical results.
Boyoung Kim, Junhong Min
Area: Object Pose Estimation
Datasets: Physics-simulated random-bin point clouds / Real robot bin-picking trials
Metrics: segmentation recall / pose accuracy / robot picking success rate
Affiliations:
Single-Shot 6DoF Pose and 3D Size Estimation for Robotic Strawberry Harvesting
A single-stage RGB detector is extended to jointly regress strawberry 6-DoF pose, 3D size, and confidence. Straw6D is automatically generated in Ignition Gazebo, with 12 equivalent rotations around the principal axis accounting for approximate symmetry. Synthetic tests run at around 60 FPS.
Lun Li, Hamidreza Kasaei
Area: Object Pose Estimation
Datasets: Straw6D / StrawDI
Metrics: 3D IoU Average Precision / rotation-translation threshold AP / inference time
Affiliations: University of Groningen
SUP-NeRF: A Streamlined Unification of Pose Estimation and NeRF for Monocular 3D Object Reconstruction
SUP-NeRF predicts and fixes car dimensions, applies three learned pose-refinement steps to projected-box states, then uses object-centric NeRF to constrain monocular scale–depth ambiguity. Errors are 7.01°/0.68m on nuScenes and 6.79°/1.06m for zero-shot KITTI. Fifty NeRF optimization steps take 1.493s without improving pose, and Waymo rotation accuracy also trails FCOS3D.
Yuliang Guo, Abhinav Kumar, Cheng Zhao, Ruoyu Wang, Xinyu Huang, Liu Ren
Area: Object Pose Estimation
Datasets: nuScenes / KITTI / Waymo Open Dataset
Metrics: rotation error / translation error / PSNR / depth error / cross-view PSNR / cross-view depth error / runtime / parameter count
Affiliations: Bosch Research North America / Bosch Center for Artificial Intelligence / Michigan State University
Learning Distributional Demonstration Spaces for Task-Specific Cross-Pose Estimation
TAX-PoseD represents the cVAE latent as a categorical spatial distribution over a point cloud, preventing multiple valid relative placements from averaging into an invalid pose. Few demonstrations allow generalization from 2-rack training to 3-rack scenes.
Jenny Wang, Octavian Donca, David Held
Area: Robot Perception & Manipulation
Datasets: NDF mug placement / PyBullet multi-rack mug hanging
Metrics: placement success rate / relative rotation error / relative translation error
Affiliations: Carnegie Mellon University
ToolEENet: Tool Affordance 6D Pose Estimation
ToolEENet estimates the pose of a tool's functional contact end instead of the whole tool. Mask R-CNN and depth extract an end-effector point cloud, score-based diffusion generates category-level 6D poses, and predicted per-axis symmetry properties handle unobservable rotations.
Yunlong Wang, Lei Zhang, Yuyang Tu, Hui Zhang, Kaixin Bai, Zhaopeng Chen, Jianwei Zhang
Area: Robot Perception & Manipulation
Datasets: TOOLEE
Metrics: 5°2cm accuracy / 5°5cm accuracy / 10°5cm accuracy / affordance segmentation mAP
Affiliations: Universität Hamburg / Agile Robots AG
U-COPE: Taking a Further Step to Universal 9D Category-level Object Pose Estimation
U-COPE treats rigid objects as articulated objects with K=1. PointNet++, part-mask feature remodulation, and within-part PPF voting jointly predict joints, segmentation, and per-part 9D poses. CAMERA25 3D75+5° is 57.3. ArtImage uses 20K point pairs by default at around 1.4–2.1 s; real ReArtMix / RobotArm results are qualitative only.
Li Zhang, Weiqing Meng, Yan Zhong, Bin Kong, Mingliang Xu, Jianming Du, Xue Wang, Rujing Wang, Liu Liu
Area: Object Pose Estimation
Datasets: CAMERA25 / ArtImage / ReArtMix / RobotArm
Metrics: 3D IoU mAP / joint rotation / translation threshold mAP / per-part rotation error / per-part translation error / per-part 3D IoU / inference time
Affiliations: Hefei Institutes of Physical Science, Chinese Academy of Sciences / University of Science and Technology of China / Anhui University / Peking University / Astribot / Hefei University of Technology
Unsupervised Learning of Category-Level 3D Pose from Object-Centric Videos
Coarse SfM meshes and DINOv2 surface features align same-category object videos into a canonical frame without pose labels. Learned neural-mesh pixel–vertex correspondences then support single-image rotation estimation by render-and-compare. CO3D alignment Acc30 is 77.0; single-image PASCAL3D+ / ObjectNet3D Acc30 is 69.2 / 52.4, at around 0.22 s per sample.
Leonhard Sommer, Artur Jesslen, Eddy Ilg, Adam Kortylewski
Area: Object Pose Estimation
Datasets: CO3D / PASCAL3D+ / ObjectNet3D
Metrics: 30° rotation accuracy / 15° rotation accuracy / mean angular error / per-sample runtime
Affiliations: University of Freiburg / Saarland University
YOLO-6D-Pose: Enhancing YOLO for Single-Stage Monocular Multi-Object 6D Pose Estimation
YOLO-6D-Pose directly regresses a continuous 6D rotation representation and translation alongside the YOLOX detection head. One forward pass detects multiple objects and estimates their poses without keypoints, PnP, or iterative refinement.
Debapriya Maji, Soyeb N. Nagori, Manu Mathew, Deepak Poddar
Area: Object Pose Estimation
Datasets: YCB-Video / LM-Occlusion
Metrics: ADD(-S) / ADD(-S) AUC
Affiliations: Texas Instruments India
Zero123-6D: Zero-Shot Novel View Synthesis for RGB Category-Level 6D Pose Estimation
Zero123-6D expands sparse reference viewpoints with pretrained diffusion. DINO cross-instance semantic matching supplies coarse poses, while offline NeUS geometry establishes 2D–3D correspondences for online optimization, enabling category-level pose without query or reference depth.
Francesco Di Felice, Alberto Remus, Stefano Gasperini, Benjamin Busam, Lionel Ott, Federico Tombari, Roland Siegwart, Carlo Alberto Avizzano
Area: Object Pose Estimation
Datasets: CO3D / Objectron
Metrics: median rotation error / rotation accuracy at 15° / rotation accuracy at 30° / online optimization time
Affiliations: Scuola Superiore Sant’Anna / Technical University of Munich / ETH Zurich / Google Zurich
ZS6D: Zero-Shot 6D Object Pose Estimation Using Vision Transformers
ZS6D retrieves the nearest CAD-rendered view with global self-supervised ViT tokens, establishes 2D–3D correspondences from local tokens, and solves pose using RANSAC-PnP. It requires no pose-specific fine-tuning and evaluates pretrained visual features' geometric generalization to unseen CAD objects.
Philipp Ausserlechner, David Haberger, Stefan Thalhammer, Jean-Baptiste Weibel, Markus Vincze
Area: Object Pose Estimation
Datasets: LM-Occlusion / YCB-Video / T-LESS / BOP
Metrics: BOP Average Recall / VSD / MSSD / MSPD
Affiliations: TU Wien / AIT Austrian Institute of Technology
CenterSnap: Single-Shot Multi-Object 3D Shape Reconstruction and Categorical 6D Pose and Size Estimation
Represents object instances by spatial centers and jointly predicts 3D shape, 6D pose, and size in one forward pass for CAD-free category-level multi-object scenes.
Muhammad Zubair Irshad, Thomas Kollar, Michael Laskey, Kevin Stone, Zsolt Kira
Area: Object Pose Estimation
Datasets: ShapeNet / NOCS / REAL275
Metrics: 3D IoU / mAP / 5° 5cm
Affiliations: Georgia Institute of Technology / Toyota Research Institute
Gen6D: Generalizable Model-Free 6-DoF Object Pose Estimation from RGB Images
Estimates CAD-free RGB 6D poses of unseen instances using posed reference images as the only object prior. Correlation-based detection is followed by view selection with global normalization and a reference-view Transformer, then iterative refinement within a 3D volume built from reference features.
Yuan Liu, Yilin Wen, Sida Peng, Cheng Lin, Xiaoxiao Long, Taku Komura, Wenping Wang
Area: Object Pose Estimation
Datasets: GenMOP / MOPED / LINEMOD
Metrics: ADD-0.1d / ADD-AUC / Prj-5
Affiliations: The University of Hong Kong / Zhejiang University / Tencent / Texas A&M University
MegaPose: 6D Pose Estimation of Novel Objects via Render & Compare
MegaPose learns category-agnostic render-and-compare from 2M synthetic images with over 20K training objects. Its coarse network tests whether 520 CAD pose hypotheses lie within the refiner's convergence basin; the refiner iteratively updates unseen-instance 6D poses using four-view RGB / depth / normal renderings. Target-specific retraining is unnecessary at test time, but metric CAD, camera intrinsics, and a target ROI are required.
Yann Labbé, Lucas Manuelli, Arsalan Mousavian, Stephen Tyree, Stan Birchfield, Jonathan Tremblay, Justin Carpentier, Mathieu Aubry, Dieter Fox, Josef Sivic
Area: Object Pose Estimation
Datasets: LM-O / T-LESS / TUD-L / IC-BIN / ITODD / HomebrewedDB / YCB-Video / ModelNet / ShapeNet / Google Scanned Objects
Metrics: BOP AR (VSD/MSSD/MSPD) / 5°/5cm recall / ADD-0.1d / Proj2D@5px
Affiliations: ENS / Inria / NVIDIA / LIGM / Ecole des Ponts / University of Washington / CIIRC CTU
MV6D: Multi-View 6D Pose Estimation on RGB-D Frames Using a Deep Point-wise Voting Network
Extends PVN3D to multiview RGB-D, fusing point clouds and image features across views to directly predict 6D poses of all objects in cluttered scenes.
Fabian Duffhauss, Tobias Demmler, Gerhard Neumann
Area: Object Pose Estimation
Datasets: YCB-Video / cluttered multi-view RGB-D datasets
Metrics: ADD / ADD-S
Affiliations: Karlsruhe Institute of Technology
PIZZA: A Powerful Image-only Zero-Shot Zero-CAD Approach to 6 DoF Tracking
Given only an initial target box in RGB video, PIZZA uses unseen-object segmentation and RAFT to maintain shared crops. A two-frame CNN or multiframe Transformer regresses inter-frame axis–angle rotation, image-plane displacement, and relative depth changes. Test objects require no CAD, depth, reference images, or object-specific training. Without a metric first-frame pose, the recovered 6D motion trajectory is relative to the initial frame, with global rotation and scale ambiguity.
Van Nguyen Nguyen, Yuming Du, Yang Xiao, Michaël Ramamonjisoa, Vincent Lepetit
Area: Object Pose Estimation
Datasets: ModelNet / ShapeNet / BOP synthetic data / Laval / MOPED
Metrics: 5°/5cm recall / ADD-0.1d / ADD-S AUC / Proj2D-5px / rotation error / translation error
Affiliations: LIGM, Ecole des Ponts, Univ Gustave Eiffel, CNRS, France
YOLOPose: Transformer-based Multi-Object 6D Pose Estimation using Keypoint Regression
Uses a Transformer for joint detection and keypoint regression, aiming to estimate detections and 6D poses for multiple objects end-to-end within a single-stage architecture.
Arash Amini, Arul Selvam Periyasamy, Sven Behnke
Area: Object Pose Estimation
Datasets: LINEMOD / Occlusion LINEMOD
Metrics: ADD / ADD-S / 2D Projection
Affiliations: University of Bonn
FFB6D: A Full Flow Bidirectional Fusion Network for 6D Pose Estimation
Uses bidirectional fusion throughout RGB and point cloud branches, then selects 3D keypoints to recover poses, strengthening appearance–geometry complementarity in RGB-D multi-object scenes.
Yisheng He, Haibin Huang, Haoqiang Fan, Qifeng Chen, Jian Sun
Area: Object Pose Estimation
Datasets: YCB-Video / LINEMOD
Metrics: ADD / ADD-S
Affiliations: Megvii Technology / HKUST
GDR-Net: Geometry-Guided Direct Regression Network for Monocular 6D Object Pose Estimation
Combines geometry-guided intermediate representations with direct pose regression, bringing geometric constraints into end-to-end monocular RGB pose estimation.
Gu Wang, Fabian Manhardt, Federico Tombari, Xiangyang Ji
Area: Object Pose Estimation
Datasets: LINEMOD / YCB-Video / T-LESS / BOP
Metrics: ADD / ADD-S / VSD / MSSD / MSPD / AR
Affiliations: Tsinghua University / Technical University of Munich / Google
CosyPose: Consistent Multi-view Multi-object 6D Pose Estimation
Models consistency across multiple objects and views, emphasizing a complete system from detection and single-view estimation to joint multiview optimization.
Yann Labbé, Justin Carpentier, Mathieu Aubry, Josef Sivic
Area: Object Pose Estimation
Datasets: YCB-Video / T-LESS / BOP
Metrics: ADD-S / VSD / MSSD / MSPD / AR
Affiliations: INRIA / École des Ponts ParisTech / CIIRC CTU
EPOS: Estimating 6D Pose of Objects with Symmetries
Represents symmetric objects with surface fragments, predicting per-pixel object probabilities, fragment probabilities, and 3D positions before running PnP-RANSAC for potentially multiple instances.
Tomas Hodan, Daniel Barath, Jiri Matas
Area: Object Pose Estimation
Datasets: T-LESS / Occlusion LINEMOD / YCB-Video / BOP
Metrics: VSD / MSSD / MSPD / AR
Affiliations: Czech Technical University in Prague
PVN3D: A Deep Point-wise 3D Keypoints Voting Network for 6DoF Pose Estimation
Performs pointwise 3D keypoint Hough voting on point clouds, then fits poses by least squares, addressing clutter and occlusion in RGB-D multi-object scenes.
Yisheng He, Wei Sun, Haibin Huang, Jianran Liu, Haoqiang Fan, Jian Sun
Area: Object Pose Estimation
Datasets: YCB-Video / LINEMOD
Metrics: ADD / ADD-S
Affiliations: Megvii Technology / Tsinghua University
DenseFusion: 6D Object Pose Estimation by Iterative Dense Fusion
Fuses RGB appearance with point cloud geometry and uses iterative refinement to improve RGB-D pose accuracy.
Chen Wang, Danfei Xu, Yuke Zhu, Roberto Martin-Martin, Cewu Lu, Li Fei-Fei, Silvio Savarese
Area: Object Pose Estimation
Datasets: YCB-Video / LINEMOD
Metrics: ADD / ADD-S
Affiliations: Stanford University / Shanghai Jiao Tong University
DPOD: 6D Pose Object Detector and Refiner
Predicts multiclass dense correspondence maps, then uses PnP/RANSAC and RGB refinement to recover multiple object 6D poses. It is a representative RGB dense-correspondence approach.
Sergey Zakharov, Ivan Shugurov, Slobodan Ilic
Area: Object Pose Estimation
Datasets: LINEMOD / Occlusion LINEMOD
Metrics: ADD / ADD-S / 2D Projection
Affiliations: Technical University of Munich / Siemens
Normalized Object Coordinate Space for Category-Level 6D Object Pose and Size Estimation
Introduces normalized object coordinate space, mapping different instances into a category-shared canonical space to estimate poses and sizes of multiple unseen instances in cluttered RGB-D scenes.
He Wang, Srinath Sridhar, Jingwei Huang, Julien Valentin, Shuran Song, Leonidas J. Guibas
Area: Object Pose Estimation
Datasets: CAMERA25 / REAL275 / NOCS
Metrics: 3D IoU / mAP / 5° 5cm
Affiliations: Stanford University / Google
PVNet: Pixel-wise Voting Network for 6DoF Pose Estimation
Localizes 2D keypoints through pixelwise direction voting, then solves pose with PnP, improving keypoint stability under occlusion.
Sida Peng, Yuan Liu, Qixing Huang, Xiaowei Zhou, Hujun Bao
Area: Object Pose Estimation
Datasets: LINEMOD / Occlusion LINEMOD
Metrics: ADD / ADD-S / 2D Projection
Affiliations: Zhejiang University
Deep Object Pose Estimation for Semantic Robotic Grasping of Household Objects
Trains an RGB keypoint / vector-field network on synthetic data to estimate 6D poses of multiple known household objects in real time for robotic grasping.
Jonathan Tremblay, Thang To, Balakumar Sundaralingam, Yu Xiang, Dieter Fox, Stan Birchfield
Area: Object Pose Estimation
Datasets: YCB-Video / YCB-Object
Metrics: ADD / ADD-S / 2D Projection
Affiliations: NVIDIA / University of Washington
PoseCNN: A Convolutional Neural Network for 6D Object Pose Estimation in Cluttered Scenes
Combines semantic segmentation, translation estimation, and rotation regression. It is a representative early deep-learning approach to 6D pose associated with the YCB-Video dataset.
Yu Xiang, Tanner Schmidt, Venkatraman Narayanan, Dieter Fox
Area: Object Pose Estimation
Datasets: YCB-Video / LINEMOD
Metrics: ADD-S / 2D Projection
Affiliations: University of Washington / NVIDIA
Real-Time Seamless Single Shot 6D Object Pose Prediction
A YOLO-style single-stage network detects objects and regresses projected 3D bounding box corners, followed by PnP, targeting real-time multi-object inference.
Bugra Tekin, Sudipta N. Sinha, Pascal Fua
Area: Object Pose Estimation
Datasets: LINEMOD / Occlusion LINEMOD
Metrics: ADD / ADD-S / 2D Projection
Affiliations: EPFL / Microsoft
SSD-6D: Making RGB-based 3D Detection and 6D Pose Estimation Great Again
Extends SSD to 3D model instance detection and 6D pose space, directly producing pose hypotheses for multiple objects from one RGB image.
Wadim Kehl, Fabian Manhardt, Federico Tombari, Slobodan Ilic, Nassir Navab
Area: Object Pose Estimation
Datasets: LINEMOD / Occlusion LINEMOD
Metrics: ADD / ADD-S / 2D Projection
Affiliations: Technical University of Munich / Siemens