CVPR 2026 · Paper Collection
CVPR 2026 object pose research continues to reduce dependence on fixed CAD, target-specific training, and dense annotations, while incorporating unified object perception, articulated parts, event cameras, and extreme egocentric scenes.
16
Papers meeting the inclusion criteria
14
Method paper
2
Data / benchmark papers
5
Generalization with fewer priors
These 16 main-conference papers form this site's core collection, selected for titles or abstracts directly addressing object 6D pose, category-level pose, articulated-object pose, object pose tracking, or related evaluation data. This is not an official CVPR task-category count.
Collection in One Sentence
A shared shift extends object pose toward generalization with fewer priors, unified object perception, continuous tracking, and real-world evaluation, beyond isolated modules for single frames, fixed objects, and individual benchmarks.
3D Gaussian Splatting, visual foundation models, VLM concepts, SE(3) equivariance, and manifold-aware regression reshape pose pipelines through geometric, semantic, and structural priors and alternative representation spaces.
Research Trends
These trends summarize research targets, input conditions, and method keywords across the included papers; individual papers may span multiple approaches.
01 · Low-prior Generalization
COG, ConceptPose, OrienPose, PoseGaussian, and PoseGAM reduce reliance on target-specific training, complete CAD, or many reference views. New-object onboarding cost is becoming as important as accuracy.
02 · Geometry as Prior
PoseGaussian uses object-level 3DGS, ConceptPose turns VLM concept vectors into 3D correspondences, and Cov2Pose places spatial second-order statistics and pose representations on an SPD manifold.
03 · Category & Articulation
ComPose, DICArt, SCAPO, and SEGPose introduce shape completion, discrete states, joint parameters, topology, and SE(3) equivariance alongside rigid-object NOCS alignment.
04 · Deployment Reality
EgoXtreme adds low light, blur, smoke, and high-speed motion to evaluation; Event6D targets high-frame-rate event cameras; KV-Tracker adapts a multiview geometry network into an online pose-tracking system.
Data Statistics
Mutually exclusive primary-theme groups show the emphasis of CVPR 2026 object pose research. Cross-theme tags remain on the paper cards.
Targets objects unseen during training, zero-shot settings, or few-reference settings, reducing dependence on CAD, category labels, and target-specific training.
Extends object and part pose through category-shared geometry, SE(3) equivariance, joint parameters, and partial observations.
Brings multiview geometry, cached visual foundation models, and event cameras into real-time object pose tracking.
Addresses extreme egocentric conditions and category-level 6D pose data production to narrow the gap between laboratory data and real deployment.
Explores more direct and complete 6D pose systems through SPD / manifold structures and unified object perception.
Interpreting the statistics: Generalization with fewer priors remains the largest individual theme, but approaches now span 3DGS, VLMs, multiview foundation models, equivariant networks, and unified perception, beyond single-image correspondence alone.
Recommended Reading
Recommendations reflect problem coverage and methodological representativeness, rather than official CVPR rankings or best-paper selections.
Unifies correspondence, confidence, and unsupervised pose estimation through optimal transport, making it relevant to few-reference generalization.
Develops object-level 3DGS into a pose prior and explicitly addresses floaters and appearance overfitting under sparse views.
Represents a shift from explicit correspondence to multiview foundation-model reasoning, incorporating large-scale synthetic object data into the generalization design.
Connects VLM concept vectors with 3D–3D correspondence, forming one of CVPR 2026's most distinctive training-free pose approaches.
Addresses novel-view synthesis through orientation-aware geometric consistency, offering a useful comparison with 3DGS approaches.
Unifies shape completion and pose reasoning to address incomplete geometry caused by partial category-level observations.
Recasts articulated-object pose as discrete-state modeling, extending category-level pose from rigid objects to parts and joint states.
Self-supervised learning of canonical geometry, part structure, and joint parameters reduces dependence on dense labels, CAD, and multiframe input for articulated pose.
Jointly optimizes object poses from multiple views into one world-frame pose, relevant to industrial multiview estimation without object-specific training.
Extends novel-object pose to high-speed event-camera tracking, with new training and evaluation data supporting deployment-oriented research.
Brings extreme lighting, motion blur, occlusion, and smoke into object pose evaluation to test generalization from laboratory settings to real wearable vision.
Moves from Euclidean regression to SPD manifold representations, providing a direct-regression alternative to correspondence / PnP pipelines.
Combines object detection, segmentation, and 6DoF pose in an object-centric transformer, representing the transition toward unified object perception systems.
Paper List
Each card gives the task setting, input modality, primary-theme tags, and a short summary. Select a card for the site's detail page, or follow the official link for the CVF Open Access abstract, PDF, and supplementary materials.
5 papers
Targets objects unseen during training, zero-shot settings, or few-reference settings, reducing dependence on CAD, category labels, and target-specific training.
Yuchen Che · Jingtu Wu · Hao Zheng · Asako Kanezaki
Setting: Single reference–query RGB-D pair; unseen objects
Input: RGB-D
Summary: Models cross-view correspondence as optimal transport with confidence-based marginal distributions, using soft correspondences and visual foundation-model semantic priors to suppress nonoverlapping regions and outliers.
Wubin Shi · Shaoyan Gai · Feipeng Da
Setting: Sparse RGB-D references + query; unseen objects
Input: RGB-D
Summary: Injects depth structure into 3DGS initialization from sparse RGB-D views, then estimates unseen-object poses through learning-guided ICP and a 3DGS render-and-compare refiner.
Jianqi Chen · Biao Zhang · Xiangjun Tang · Peter Wonka
Setting: Query RGB + multiple templates; objects unseen during training
Input: RGB
Summary: Takes a query and multiple templates, directly predicting object pose through explicit point geometry and a geometric representation network without manually constructed query–template correspondences. Includes synthetic data covering more than 190K objects.
Liming Kuang · Yordanka Velikova · Mahdi Saleh · Jan-Nico Zaech · Danda Pani Paudel · Benjamin Busam
Setting: Zero-shot relative pose; unseen objects
Input: RGB-D / 3D
Summary: Uses VLM saliency to generate open-vocabulary 3D concept maps, then establishes cross-view 3D–3D correspondences to estimate 6DoF relative pose without object-specific or dataset-specific training.
Yating Liu · Zhaoshuai Qi · Yang Zou · Yongnan Yang · Shizhou Zhang · Yanning Zhang
Setting: Single-image unseen-object pose without CAD
Input: RGB
Summary: Explicitly injects object orientation into single-image unseen-object pose estimation and constrains novel-view geometry with an orientation consistency loss, reducing the effect of distorted and blurred templates on matching.
4 papers
Extends object and part pose through category-shared geometry, SE(3) equivariance, joint parameters, and partial observations.
Huan Ren · Yihan Chen · Chuxin Wang · Nailong Liu · Wenfei Yang · Tianzhu Zhang
Setting: Category-level pose + size from partial point clouds
Input: RGB-D / point cloud
Summary: Recovers complete shapes through keypoint-based progressive completion, then uses geometric relation encoding and a geometric relation consistency loss to constrain global alignment between keypoints and NOCS coordinates.
Li Zhang · Mingyu Mei · Ailing Wang · Xianhui Meng · Yan Zhong · Xinyuan Song · Liu Liu · Rujing Wang · Zaixing He · Cewu Lu
Setting: Category-level articulated-object and part pose
Input: 3D / point cloud
Summary: Models part poses and discrete states of category-level articulated targets, aiming to recover structured object and part poses under both category and articulation-state variation.
Can Zhang · Gim Hee Lee
Setting: Single RGB-D / 3D observation; category-level articulated targets
Input: RGB-D / point cloud
Summary: First aligns different instances to a shared canonical space through an SE(3)-equivariant autoencoder, then estimates parts, joint axes, pivots, and articulation states through joint-aware blend skinning.
Sheng Yu · Di-Hua Zhai · Yuanqing Xia
Setting: Category-level object pose
Input: 3D / point cloud
Summary: Combines SE(3)-equivariant representations with geometric and topological information to strengthen modeling of shape structure and spatial transformations in category-level pose prediction.
3 papers
Brings multiview geometry, cached visual foundation models, and event cameras into real-time object pose tracking.
Anna Šárová Mikeštíková · Médéric Fourmy · Martin Cifka · Josef Sivic · Vladimir Petrik
Setting: Multiple extrinsically calibrated RGB views; known models, potentially unseen targets
Input: Multi-view RGB
Summary: Uses multiple extrinsically calibrated RGB images and jointly minimizes rendered-versus-observed object feature discrepancies in a frozen feature space, without target-specific training or symmetry annotations.
Jae-Young Kang · Hoonhee Cho · Taeyeop Lee · Minjun Kang · Bowen Wen · Youngho Kim · Kuk-Jin Yoon
Setting: Unseen-object pose tracking in high-speed dynamic scenes
Input: Event + depth
Summary: Reconstructs intensity/depth at arbitrary timestamps from event streams and depth frames for novel-object 6D pose tracking. Reports over 120 FPS and includes synthetic training data plus real and simulated evaluation sets.
Marwan Taher · Ignacio Alzugaray · Kirill Mazur · Xin Kong · Andrew Davison
Setting: Online object pose tracking and reconstruction from monocular RGB video
Input: Monocular RGB
Summary: Caches global attention from a multiview visual geometry network as a key–value scene representation for online object and scene tracking, improving runtime and reducing drift without retraining.
2 papers
Addresses extreme egocentric conditions and category-level 6D pose data production to narrow the gap between laboratory data and real deployment.
Taegyoon Yoon · Yegyu Han · Seojin Ji · Jaewoo Park · Sojeong Kim · Taein Kwon · Hyung-Sin Kim
Setting: Egocentric 6D pose in industrial maintenance, sports, and emergency response
Input: Egocentric RGB
Summary: Introduces a large-scale 6D pose dataset for smart-glasses-style egocentric vision, covering industrial maintenance, sports, and emergency response, and shows substantial generalization degradation under low light and extreme conditions.
Duret Guillaume · Danylo Mazurak · Florence Zara · Jan Peters · Liming Chen
Setting: Category-level 6D pose data generation and robotic grasping
Input: 3D assets / synthetic data
Summary: Explores rapid generation of aligned 3D assets from scratch for category-level 6D pose and robotic grasping, addressing the difficulty of scaling high-quality 3D training data production.
2 papers
Explores more direct and complete 6D pose systems through SPD / manifold structures and unified object perception.
Nassim Ali Ousalah · Peyman Rostami · Vincent Gaudillière · Emmanuel Koumandakis · Anis Kacem · Enjie Ghorbel · Djamila Aouada
Setting: Direct 6DoF object pose regression from a single RGB image
Input: RGB
Summary: Builds symmetric positive-definite matrix features from spatial covariance and parameterizes rotation and translation through Cholesky decomposition, directly regressing object pose with a Riemannian geometry-aware head.
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
Setting: Unified 3D object perception with joint detection, segmentation, and pose prediction
Input: RGB / multi-view geometry
Summary: Unifies detection, segmentation, and 6DoF pose using 2D and 3D object-centric transformers, improving zero-shot cross-domain object perception through continuous refinement and multiview information.
Scope & Sources
Inclusion: CVPR 2026 main-conference papers whose titles or abstracts directly address object 6D pose, category-level pose, articulated-object pose, object pose tracking, or datasets and data generation directly supporting these tasks.
Excluded: human pose, relative camera / scene pose, visual tracking without object pose output, 3D reconstruction alone, robot grasp pose, and CVPR Findings or workshop papers. These neighboring works can be added later in a separate related-directions section.
Primary sources: CVPR 2026 Papers, CVF Open Access Repository and each paper's official page. Summaries are editorial explanations and do not replace the original papers.
