01 · Generalization
Unseen objects become the central problem
Co-op, Any6D, RefPose, One2Any, Pos3R, and UNOPose reduce dependence on fixed categories, complete CAD, or many reference viewpoints, extending generalization to arbitrary new objects.
CVPR 2025 · Annual Review
Object Pose Estimation: Annual Review
CVPR 2025 object pose research is shifting toward generalizable estimation for objects unseen during training, with fewer priors and references. This page organizes main-conference papers under consistent inclusion criteria and distinguishes official presentation designations from editorial analysis.
17
Papers meeting the inclusion criteria
2
Official Highlight
17
Official Poster pages
0
Orals within core scope
These 17 papers reflect this site's inclusion criteria, rather than an official CVPR task-category count. Official schedules and paper pages are the primary sources for verifying titles and Poster / Highlight designations.
View the official CVPR 2025 schedule →Year in One Sentence
From model availability to generalization with fewer priors
A common question this year is whether a system can still produce usable 6D poses without complete target CAD, with only one reference, across nonoverlapping viewpoints, under occlusion, or outside the training domain.
Reference-image encoding, 3D reconstruction foundation models, category-shared representations, test-time adaptation, uncertainty, and modular refinement therefore form intersecting research themes.
Research Trends
Research trends
These trends summarize the main problems and methodological keywords of the 17 included papers; a paper may contribute to multiple trends.
02 · Reference to Geometry
A single reference image becomes an implicit 3D prior
One2Any encodes reference objects with ROPE / ROC; Any6D jointly aligns pose and scale; Pos3R uses a 3D foundation model; UA-Pose explicitly models partial observations as geometry that can be completed, with uncertainty.
03 · Category-level
Category-level methods expand from NOCS to structure and context
GIVEPose reexamines within-category variation in NOCS, SpotPose emphasizes shape-sensitive correspondence and outlier removal, and GCE-Pose and Global Stereo Consistency introduce global shape or cross-view context.
04 · Deployment Gap
Domain gaps, occlusion, and test-time correction constrain systems
CRISP and ONDA-Pose explicitly address domain gaps, pseudo-labels, occlusion, and test-time adaptation after training. SCFlow2 designs refinement as a modular component.
Data Statistics
Statistics and classification
Counts use mutually exclusive primary themes assigned by this site to show the year's emphasis; cross-theme tags are retained in the paper list below.
Unseen objects and few reference views
8Generalizable estimation for objects unseen during training, starting from CAD, a single reference image, or a few RGB-D observations.
Category-level and category-agnostic estimation
6Improves pose and scale estimation across instances through category-shared coordinates, structural priors, keypoints, and multiview geometry.
Domain adaptation and test-time learning
2Treats synthetic-to-real transfer, occlusion, and deployment-domain gaps as central problems, exploring model correction at test time.
Modular pose refinement
1Focuses on RGB-D refinement, scene flow, and geometric constraints after initial pose estimation.
Interpreting the statistics: Unseen objects and few-reference estimation form the largest theme, followed by category-level / category-agnostic estimation. The emphasis in 2025 is shifting toward new-object onboarding cost, cross-domain stability, and deployability.
Recommended Reading
Recommended papers
Reading priorities combine official Highlight designations, problem coverage, and representativeness for subsequent research directions; they are not an official CVPR ranking.
Representative approach
Any6D
Useful for understanding the emerging route from a single reference through generated geometry to render-and-compare.
Annual focus Official Highlight
CRISP
Connects pose, shape, geometric correction, and test-time self-training in a complete system, making it one of this year's system papers most worth reading closely.
Annual focus Official Highlight
CAP-Net
Combines articulated objects, category-level pose/size, and a new dataset, connecting pose estimation with robotic manipulation.
Representative approach
One2Any
A representative shift from dependence on multiview data or CAD to a single reference image; useful alongside Any6D and UNOPose.
Representative approach
Pos3R
Shows how a 3D reconstruction foundation model can directly supply geometric priors to a pose pipeline.
Representative approach
SpotPose
Useful for understanding feature quality and outlier rejection as key bottlenecks in category-level correspondence pipelines.
About Oral presentations: For the core object pose papers included here, official paper pages and schedules currently verify Poster presentations; CRISP and CAP-Net also carry official Highlight labels. Related visual localization, camera pose, human pose, and 3D reconstruction papers are excluded from this count.
Paper List
Paper list and summaries
Select a card for the site's paper detail page, including coarse and deep notes and original figures and tables. The official CVPR page remains linked at the lower right.
8 papers
Unseen objects and few reference views
Generalizable estimation for objects unseen during training, starting from CAD, a single reference image, or a few RGB-D observations.
CVPR 2025 Poster
Co-op: Correspondence-based Novel Object Pose Estimation
Sungphill Moon · Hyeontae Son · Dongcheol Hur · Sangwook Kim
novel objectCAD modelsemi-dense correspondencedifferentiable PnP
Setting: Single RGB image + target CAD; objects unseen during training
Input: RGB
Summary: Establishes semidense correspondences from a few rendered templates, then combines patch classification, offset regression, and probabilistic flow for pose refinement without target-object fine-tuning.
CVPR 2025 Poster
Any6D: Model-free 6D Pose Estimation of Novel Object
Taeyeop Lee · Bowen Wen · Minjun Kang · Gyuree Kang · In So Kweon · Kuk-Jin Yoon
one referenceRGB-Dmodel-freepose + sizerender-and-compare
Setting: Single RGB-D anchor + query; unseen objects
Input: RGB-D
Summary: Estimates an unseen object's pose and size from one RGB-D anchor, using joint object alignment and render-and-compare to handle occlusion, nonoverlapping views, and environmental changes.
CVPR 2025 Poster
iG-6DoF: Model-free 6DoF Pose Estimation for Unseen Object via Iterative 3D Gaussian Splatting
Tuo Cao · Fei Luo · Jiongming Qin · Yu Jiang · Yusen Wang · Chunxia Xiao
unseen object3D Gaussian Splattingrotation-equivariant featureiterative refinement
Setting: Model-free unseen-object pose
Input: RGB
Summary: Generates initial pose hypotheses using multiscale augmentation and rotation-equivariant features, then performs coarse-to-fine refinement through iterative 3D Gaussian Splatting render-and-compare, targeting large displacements and weak texture.
CVPR 2025 Poster
RefPose: Leveraging Reference Geometric Correspondences for Accurate 6D Pose Estimation of Unseen Objects
Jaeguk Kim · Jaewoo Park · Keuntek Lee · Nam Ik Cho
unseen objectreference imagegeometric correspondencerefinement
Setting: Monocular RGB + reference image; unseen objects
Input: RGB
Summary: Uses reference images to guide geometric correspondence: obtains an initial pose from template renderings and correspondences, then refines it for accurate monocular estimation of objects unseen during training.
CVPR 2025 Poster
One2Any: One-Reference 6D Pose Estimation for Any Object
Mengya Liu · Siyuan Li · Ajad Chhatkuli · Prune Truong · Luc Van Gool · Federico Tombari
one referenceRGB-DROPEROCcategory-agnostic
Setting: Single reference–query RGB-D pair; relative pose of arbitrary objects
Input: RGB-D
Summary: Encodes object shape, pose, and texture through Reference Object Pose Embedding, then predicts Reference Object Coordinate for single-reference pose estimation without CAD, multiple views, or category priors.
CVPR 2025 Poster
Pos3R: 6D Pose Estimation for Unseen Objects Made Easy
Weijian Deng · Dylan Campbell · Chunyi Sun · Jiahao Zhang · Shubham Kanitkar · Matthew Shaffer · Stephen Gould
unseen objectsingle RGB3D foundation modelBOP
Setting: Single RGB image; unseen objects; no additional training
Input: RGB
Summary: Uses a 3D reconstruction foundation model to produce correspondences with greater 3D consistency, estimating unseen-object pose without additional pose training and optionally adding render-and-compare refinement.
CVPR 2025 Poster
UNOPose: Unseen Object Pose Estimation with an Unposed RGB-D Reference Image
Xingyu Liu · Gu Wang · Ruida Zhang · Chenyangguang Zhang · Federico Tombari · Xiangyang Ji
unseen objectone referenceRGB-DSE(3)-invariant frame
Setting: Single unposed RGB-D reference; unseen objects
Input: RGB-D
Summary: Simplifies reference acquisition from a posed anchor to one unposed RGB-D image, constructs an SE(3)-invariant reference frame, and reweights correspondences by overlap probability.
CVPR 2025 Poster
UA-Pose: Uncertainty-Aware 6D Object Pose Estimation and Online Object Completion with Partial References
Ming-Feng Li · Xin Yang · Fu-En Wang · Hritam Basak · Yuyin Sun · Shreekant Gayaka · Min Sun · Cheng-Hao Kuo
partial referenceuncertaintyonline completionnovel object
Setting: Partial RGB-D reference or a single 2D image
Input: RGB / RGB-D
Summary: Explicitly separates observed and unobserved regions of partial 3D models, using uncertainty for pose confidence and online object completion under incomplete reference information.
6 papers
Category-level and category-agnostic estimation
Improves pose and scale estimation across instances through category-shared coordinates, structural priors, keypoints, and multiview geometry.
CVPR 2025 Poster Highlight
CAP-Net: A Unified Network for 6D Pose and Size Estimation of Categorical Articulated Parts from a Single RGB-D Image
Jingshun Huang · Haitao Lin · Tianyu Wang · Yanwei Fu · Xiangyang Xue · Yi Zhu
Highlightarticulated partscategory-levelRGBD-Art
Setting: Single RGB-D image; category-level articulated parts
Input: RGB-D
Summary: Introduces a single-stage RGB-D network for category-level articulated parts, jointly predicting pointwise classes, centroid offsets, and NPCS, and releases RGBD-Art for sim-to-real evaluation.
CVPR 2025 Poster
Leveraging Global Stereo Consistency for Category-Level Shape and 6D Pose Estimation from Stereo Images
Junning Qiu · Minglei Lu · Fei Wang · Yu Guo · Yonggen Ling
category-levelstereoshape estimationglobal consistency
Setting: Stereo images; category-level shape + 6D pose
Input: Stereo RGB
Summary: Represents a shared shape through an intra-category occupancy field and constrains shape, pose, and other variables through global stereo consistency, reducing parameter coupling and shape ambiguity in stereo settings.
CVPR 2025 Poster
Recurrent Feature Mining and Keypoint Mixup Padding for Category-Agnostic Pose Estimation
Junjie Chen · Weilong Chen · Yifan Zuo · Yuming Fang
category-agnosticfew-shot supportkeypoint localizationfeature mining
Setting: Few support images + query; category-agnostic keypoint pose
Input: RGB
Summary: Targets support–query pose estimation for arbitrary novel categories, emphasizing fine-grained, structure-aware features and keypoint mixup padding for better pixel-level keypoint localization.
CVPR 2025 Poster
GIVEPose: Gradual Intra-class Variation Elimination for RGB-based Category-Level Object Pose Estimation
Ziqin Huang · Gu Wang · Chenyangguang Zhang · Ruida Zhang · Xiu Li · Xiangyang Ji
category-levelRGB-onlyNOCSintra-class variation
Setting: RGB category-level pose
Input: RGB
Summary: Argues that NOCS maps retain redundant instance variation in category-level pose, introducing an IVFC map that complements NOCS to progressively remove within-category variation.
CVPR 2025 Poster
Rethinking Correspondence-based Category-Level Object Pose Estimation
Huan Ren · Wenfei Yang · Shifeng Zhang · Tianzhu Zhang
category-levelcorrespondencepose-invariant featureoutlier removal
Setting: Category-level correspondence + pose fitting
Input: RGB-D / geometric input
Summary: Reexamines correspondence-based category-level pose and combines shape-sensitive, pose-invariant features with explicit outlier-correspondence removal in SpotPose.
CVPR 2025 Poster
GCE-Pose: Global Context Enhancement for Category-level Object Pose Estimation
Weihang Li · Hongli Xu · Junwen Huang · HyunJun Jung · Peter K. T. Yu · Nassir Navab · Benjamin Busam
category-levelpartial visibilityshape reconstructionglobal context
Setting: Partially visible RGB-D category-level pose
Input: RGB-D
Summary: First completes an instance's global semantic shape using category prototypes, then aggregates observed and reconstructed context to address incomplete category-level features under partial visibility.
2 papers
Domain adaptation and test-time learning
Treats synthetic-to-real transfer, occlusion, and deployment-domain gaps as central problems, exploring model correction at test time.
CVPR 2025 Poster Highlight
CRISP: Object Pose and Shape Estimation with Test-Time Adaptation
Jingnan Shi · Rajat Talak · Harry Zhang · David Jin · Luca Carlone
Highlightcategory-agnosticpose + shapetest-time adaptation
Setting: RGB-D; category-agnostic pose and shape
Input: RGB-D
Summary: Estimates pose through FiLM-conditioned implicit shape reconstruction and pose-normalized points, then uses an active shape model for correct-and-certify and test-time self-training to reduce the domain gap.
CVPR 2025 Poster
ONDA-Pose: Occlusion-Aware Neural Domain Adaptation for Self-Supervised 6D Object Pose Estimation
Tao Tan · Qiulei Dong
self-superviseddomain adaptationocclusion-awaresynthetic-to-real
Setting: Self-supervised 6D pose for synthetic-to-real transfer
Input: RGB
Summary: Develops occlusion-aware neural domain adaptation to address the synthetic-to-real image gap and real occlusions that image translation cannot adequately cover.
1 papers
Modular pose refinement
Focuses on RGB-D refinement, scene flow, and geometric constraints after initial pose estimation.
CVPR 2025 Poster
SCFlow2: Plug-and-Play Object Pose Refiner with Shape-Constraint Scene Flow
Qingyuan Wang · Rui Song · Jiaojiao Li · Kerui Cheng · David Ferstl · Yinlin Hu
pose refinementRGB-Dscene flowshape constraint
Setting: Modular RGB-D pose refinement
Input: RGB-D
Summary: Adds depth constraints to SCFlow's iterative RGB refinement, formulating the additional depth information as 3D scene-flow regularization to reduce correspondence noise and support novel objects.
Scope & Sources
Counting rules and scope
Inclusion: CVPR 2025 main-conference papers whose titles or abstracts directly address object 6D pose, category-level pose, category-agnostic pose, unseen-object pose, or object pose refinement.
Excluded: human pose, relative camera pose, scene reconstruction alone, tracking alone, workshop papers, and papers that use pose only in experiments without treating it as the central problem.
Primary sources: CVPR 2025 Papers, CVPR 2025 Schedule and each paper's official Poster / Open Access page. Summaries are editorial explanations and do not replace the original papers.