3DV 2026 · Paper Collection
Direct 6D pose papers are relatively few at 3DV 2026, but surrounding representations are expanding: a few templates onboard new objects, partial observations map to canonical frames, and hand–object and articulated-object methods jointly reconstruct pose, geometry, motion, and contact.
7
Papers in this collection
3
Direct pose / alignment
4
Interaction and articulated objects
1
Included Orals
The official 3DV directory lists 177 main-conference papers, including 22 Orals. This collection includes direct object pose, CAD alignment, and canonicalization, plus hand–object and articulated-object work explicitly recovering object transformations, joint parameters, or part motion. Camera pose alone, generic SfM, 3D detection, and rendering alone are excluded.
Collection in One Sentence
NeMO, CAOA, and C3PO remain central to pose, but go beyond single-frame regression: template memory, completion, and correspondence / registration rethink object priors, incomplete observations, and canonical frames.
The other four papers place objects in real interaction, using hands as geometric anchors and including joints and part motion in their outputs. For robotics, these approaches support obtaining manipulable object models and stably tracking their states, even when BOP scores are not the endpoint.
Research Trends
These themes summarize the 7 included papers by task conditions and method structure, rather than the entire 3DV research distribution.
01 · Object Onboarding
NeMO builds semantic and geometric object memory from a few RGB templates; CAOA completes geometry before fitting CAD to real scans. Both shorten the path from an unknown object to one that can be localized.
02 · Partial Geometry
C3PO handles local video segments through generalizable correspondences and pairwise registration. It highlights the importance of consistent canonical frames across instances and observations, alongside rotation / translation prediction.
03 · Interaction State
Hand-held reconstruction, Follow My Hold, GaussianArt, and iTACO include occlusion, rigid alignment, joint parameters, and part motion in object state, supporting manipulation and digital twins.
Data Statistics
Each paper belongs to one group by its main problem setting; card tags retain cross-area attributes.
Direct 6DoF / 9DoF object pose, or alignment of partial observations to a common canonical frame.
Uses hand geometry, contact, and temporal observations to recover object shape and relative transformations under severe occlusion.
Unifies part geometry, joint parameters, and motion decomposition to model subsequent part-level pose and interaction.
Interpreting the statistics: The 3 direct pose / alignment papers form the strict core. Another 4 recover object geometry, motion, and interaction states. They are distinguished from BOP-style 6D pose comparisons but help explain 3DV 2026's broader treatment of object state.
Recommended Reading
Recommendations reflect problem coverage and methodological representativeness; Oral denotes official 3DV scheduling rather than editorial assessment.
The only included Oral directly targeting 6DoF object pose. Compresses a few templates into object-centric memory for fast onboarding, covering detection, segmentation, and BOP pose.
Combines point completion, symmetry, and 9DoF alignment in one evaluable pipeline, closely matching CAD-to-object pose fitting in real scans.
Recasts category-level pose as partial-view canonicalization, using cross-category correspondence features and global registration to address unknown categories.
Treats simultaneous camera motion, object motion, and hand occlusion as default conditions, targeting joint / part recovery for interactive digital twins.
Paper List
Each card opens the site's paper detail page; official 3DV paper and PDF links remain at the lower right.
3 papers
Direct 6DoF / 9DoF object pose, or alignment of partial observations to a common canonical frame.
Sebastian Jung · Leonard Klüpfel · Rudolph Triebel · Maximilian Durner
Setting: Detection, segmentation, and 6DoF pose of objects unseen during training from a few RGB template views
Input: few-shot RGB templates + query RGB
Summary: NeMO's encoder converts a few RGB templates into a sparse object-like point cloud containing semantics and geometry. Its decoder combines object encodings with query images for dense perception and unseen-object 6DoF pose, without target-specific retraining or camera-specific parameters.
Hiranya Garbha Kumar · Minhas Kamal · Balakrishnan Prabhakaran
Setting: CAD alignment to incomplete, noisy real objects in indoor RGB-D scans
Input: indoor RGB-D scan / object point cloud + CAD
Summary: CAOA repairs missing scan geometry through semantic- and context-aware point-cloud completion, then aligns CAD through symmetry-aware relative pose estimation. Releases S2C-Completion annotations derived from Scan2CAD and reports a 17% accuracy improvement over previous methods on Scan2CAD.
Yu Chi · Leonhard Sommer · Olaf Dünkel · Dominik Muhle · Daniel Cremers · Christian Theobalt · Adam Kortylewski
Setting: Category-level canonical frames from short videos / partial point clouds with incomplete coverage
Input: multi-view images → SfM point clouds
Summary: C3PO reconstructs short videos into partial point clouds with 2D features, learns cross-category correspondence features, and aligns multiple local observations into a globally consistent canonical frame through efficient pairwise registration.
2 papers
Uses hand geometry, contact, and temporal observations to recover object shape and relative transformations under severe occlusion.
Jane Wu · Georgios Pavlakos · Georgia Gkioxari · Jitendra Malik
Setting: Geometry reconstruction and temporal rigid alignment of hand-occluded objects in internet images / videos
Input: monocular RGB image / video + inferred 3D hand
Summary: MCC-Hand-Object jointly recovers hand and object geometry from a single image and an inferred 3D hand. RAR uses VLM-assisted retrieval to find a matching 3D object model and rigidly aligns it with images and MCC-HO observations consistently over time.
Ayce Idil Aytekin · Helge Rhodin · Rishabh Dabral · Christian Theobalt
Setting: Occluded-object reconstruction from monocular RGB through hand–object geometry, contact, and keypoints
Input: monocular RGB + hand-object geometric cues
Summary: Jointly optimizes hand and reconstructed-object transformations during diffusion, guided by normals / depth, silhouettes, 2D keypoint reprojection, and SDF contact constraints for more physically consistent hand–object reconstruction under occlusion.
2 papers
Unifies part geometry, joint parameters, and motion decomposition to model subsequent part-level pose and interaction.
Licheng Shen · Saining Zhang · Honghan Li · Peilin Yang · Zihao Huang · Zongzheng Zhang · Hao Zhao
Setting: Joint part-geometry and motion modeling for articulated objects
Input: articulated-object visual observations
Summary: GaussianArt represents geometry and motion together through articulated 3D Gaussians, avoiding fragile separate-state reconstruction followed by alignment. Introduces MPArt-90 with 90 articulated objects spanning 20 categories and up to 20 parts for evaluating part-level reconstruction and motion estimation.
Weikun Peng · Jun Lv · Cewu Lu · Manolis Savva
Setting: Joint-parameter and movable-part recovery from casually captured RGB-D interaction videos
Input: casually captured dynamic RGB-D video
Summary: iTACO uses a coarse-to-fine pipeline to infer joint parameters and movable-part segmentation from dynamic RGB-D video captured with handheld cameras. Introduces a dataset of 784 videos, 284 objects, and 11 categories, approximately 20 times larger than previous comparable data.
Scope & Sources
Inclusion: officially accepted 3DV 2026 main-conference papers whose titles or abstracts directly address object 6DoF / 9DoF pose, object–CAD alignment, category-level canonicalization, or hand–object / articulated-object work explicitly outputting object transformations, joint parameters, or part motion.
Excluded: camera pose / SfM / SLAM alone, 3D object detection, human pose, generic 3D reconstruction or novel-view synthesis alone, and work with only downstream links to object pose. Hand–object and articulated-object papers are grouped as reconstruction and modeling to distinguish them from direct pose benchmarks.
Primary sources: Official 3DV 2026 paper directory and linked official OpenReview paper pages. Summaries and trends are editorial work and do not replace the original papers.
