Method Topic
Multi-object Pose Estimation
This collection focuses on 6D pose methods designed for multiple objects or instances, scene-level consistency, joint multiview optimization, or temporal fusion. Individual entries can grow into detailed paper pages.
62
Collected papers
5
Task settings
2017-2026
Publication span
Inclusion prioritizes the method's objective: estimating multiple poses in one scene or explicitly using inter-object, multiview, or temporal relationships for joint estimation. Methods merely evaluated object by object on multi-object datasets can be listed separately as related approaches.
By Setting
Browse by input / output setting
Methods within one setting often share requirements, failure modes, and experimental protocols, making comparisons easier as detailed paper notes are added.
RGB → multi-instance 6D poses
Monocular RGB: joint detection and pose
These methods jointly model detection, keypoint / correspondence prediction, and pose solving. They suit RGB-only settings that must process multiple known objects in a scene at once.
SSD-6D
Extends SSD to 3D model instance detection and 6D pose space, directly producing pose hypotheses for multiple objects from one RGB image.
Single Shot 6D
A YOLO-style single-stage network detects objects and regresses projected 3D bounding box corners, followed by PnP, targeting real-time multi-object inference.
DOPE
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.
PoseCNN
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.
DPOD
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.
EPOS
Represents symmetric objects with surface fragments, predicting per-pixel object probabilities, fragment probabilities, and 3D positions before running PnP-RANSAC for potentially multiple instances.
YOLOPose
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.
RGB-D / point cloud → per-object 6D poses
RGB-D: point cloud geometry and keypoint voting
Depth makes occlusion, scale, and geometric constraints more accessible in multi-object scenes. Common approaches vote for 3D keypoints from point clouds or fused RGB-D features, then recover each instance's pose.
PVN3D
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.
FFB6D
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.
RGB-D + category prior → pose + size
Category-level: multiple instances without exact CAD
These methods handle unseen objects within known categories rather than fixed known CAD instances, simultaneously estimating poses and scales for multiple instances. Some also predict shape jointly.
NOCS
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.
CenterSnap
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.
multi-view RGB/RGB-D → consistent scene poses
Multiple views: scene consistency and joint optimization
Multiple views substantially reduce occlusion and single-view ambiguity. The central challenge shifts from local image predictions to cross-view matching and consistent optimization of camera and object poses.
CosyPose
Models consistency across multiple objects and views, emphasizing a complete system from detection and single-view estimation to joint multiview optimization.
MV6D
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.
video sequence → temporally fused poses
Video: temporal evidence for multiple objects
In dynamic scenes, occlusion and motion can destabilize single-frame predictions. Temporal methods aggregate detection and pose evidence across frames for more stable multi-object outputs.
Timeline
Browse by publication date
Follow the progression from single-stage RGB detection to RGB-D point cloud voting, category-level multiple instances, multiview optimization, and temporal video fusion.
2017
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.
2018
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.
2018
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.
2018
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.
2019
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.
2019
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.
2020
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.
2020
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.
2020
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.
2021
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.
2022
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.
2022
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.
2022
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.
2024
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.
2024
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.
2024
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°.
2024
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.
2024
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.
2024
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.
2024
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.
2024
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.
2024
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.
2024
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.
2024
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.
2024
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.
2024
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.
2024
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.
2024
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.
2024
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.
2025
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.
2025
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.
2025
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.
2025
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.
2025
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.
2025
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.
2025
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.
2025
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.
2025
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.
2025
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.
2025
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.
2025
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.
2025
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.
2025
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.
2025
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.
2026
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.
2026
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.
2026
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.
2026
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.
2026
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.
2026
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.
2026
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.
2026
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.
2026
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.
2026
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.
2026
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.
2026
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.
2026
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.
2026
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.
2026
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.
2026
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.
2026
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.
2026
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.