AAAI 2026 · Paper Collection
AAAI 2026 has fewer concentrated object-pose papers than CVPR, but covers important branches from few-reference unseen-object estimation and category-level articulated-part tracking to hand–object interaction and 6DoF model security.
5
Papers meeting the inclusion criteria
4
Method paper
1
Security research
1
Few-reference generalization
These 5 papers form this site's AAAI-26 core collection, selected for titles or abstracts directly addressing object 6D pose, category-level pose, articulated-object pose tracking, hand–object pose, or 6DoF pose-model security. This is not a count of all AAAI computer-vision papers.
Collection in One Sentence
AAAI 2026 pose papers bridge Computer Vision, Robotics, and AI Safety. They reduce dependence on CAD and multiview modeling for unseen objects while extending pose into articulated parts, hand–object contact, and security-attack settings.
Coordinate tokens, autoregressive generation, SE(3) manifolds, physical cues, and backdoor triggers address global correspondence, structural constraints, interaction consistency, and deployment reliability.
Research Trends
Trends summarize task settings and method keywords across the included papers; individual papers may touch multiple directions.
01 · Minimal Reference
CoordAR discretizes reference–query 3D correspondences into token maps generated by an autoregressive decoder, reflecting the need for global consistency and uncertainty modeling in few-reference unseen-object pose.
02 · Structured Interaction
PPF-Tracker uses joint axes and SE(3) structure for articulated tracking. VPHO combines visual and physical cues to reduce hand–object interpenetration and missing contact, extending alignment toward interaction feasibility.
03 · Reliability by Design
MonoVPR addresses geometric degradation in distant monocular vehicle pose. 6DAttack shows attacker control of continuous rotation and translation outputs, motivating evaluation of robustness and security together.
Data Statistics
Each paper belongs to one group by its main problem setting; overlapping keywords remain on the paper cards.
Studies stable geometric correspondence from very few references for new objects without CAD or complete 3D models.
Combines category-level objects, joint axes, and multiframe pose tracking with SE(3) structure and motion constraints.
Recovers object geometry and 3D pose together in distant scenes such as autonomous driving, reducing monocular depth ambiguity.
Uses visual and physical cues to constrain hand and object 3D poses jointly, improving contact and interaction plausibility.
Introduces security issues such as backdoor attacks into 6DoF pose estimation to examine failure boundaries under attack.
Interpreting the statistics: AAAI-26 includes few direct object-pose methods, but spans representative problems: novel-object geometry, articulated objects, hand–object interaction, monocular vehicle shapes, and model security, with a scope broader than one benchmark.
Recommended Reading
Recommendations reflect problem coverage and methodological representativeness, rather than official AAAI rankings or best-paper selections.
Discretizes 3D correspondence into autoregressively generated coordinate maps, representing a shift toward probabilistic global modeling at AAAI 2026.
Combines point-pair features, SE(3) invariance, and joint-axis constraints for category-level articulated pose tracking.
Demonstrates joint pose-and-shape modeling for distant monocular scenes, emphasizing dynamic fusion of object-centric geometry and scene-centric semantics.
Adds contact, interpenetration, and physical-feasibility constraints to visual alignment, relevant to embodied perception and interactive pose estimation.
Extends pose-model security to precise attacker control of rotation and translation, highlighting risks associated with continuous pose outputs in deployment.
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 abstracts, PDFs, and citation information on AAAI OJS.
1 papers
Studies stable geometric correspondence from very few references for new objects without CAD or complete 3D models.
Dexin Zuo · Ang Li · Wei Wang · Wenxian Yu · Danping Zou
Setting: Single reference–query pair for unseen objects without complete 3D models
Input: RGB-D / 3D correspondence
Summary: Encodes reference–query 3D–3D correspondences as discrete token maps and predicts them through an autoregressive Transformer combining RGB appearance and coordinate cues, modeling uncertainty under symmetry and occlusion.
1 papers
Combines category-level objects, joint axes, and multiframe pose tracking with SE(3) structure and motion constraints.
Xianhui Meng · Yukang Huo · Li Zhang · Liu Liu · Haonan Jiang · Yan Zhong · Pingrui Zhang · Cewu Lu · Jun Liu
Setting: Multiframe 6D pose tracking of category-level articulated objects
Input: RGB-D / point cloud
Summary: Performs quasi-canonicalization in the SE(3) Lie group, predicts pose-voting parameters through Point Pair Features, and applies shared kinematic constraints to different parts using joint-axis semantics.
1 papers
Recovers object geometry and 3D pose together in distant scenes such as autonomous driving, reducing monocular depth ambiguity.
Wei Li · Long Ji · Ying Wang · Xiao Wu · Zhaoquan Yuan · Penglin Dai
Setting: Distant monocular vehicle 3D pose and shape recovery in autonomous driving
Input: Monocular RGB
Summary: Fuses multiscale object and scene information through hierarchical dual-context attention, then progressively adjusts template shapes with a bounded iterative mesh refiner while recovering more stable vehicle 3D poses.
1 papers
Uses visual and physical cues to constrain hand and object 3D poses jointly, improving contact and interaction plausibility.
Jun Zhou · Chi Xu · Kaifeng Tang · Yuting Ge · Tingrui Guo · Li Cheng
Setting: Hand and object 3D poses in single-RGB hand–object interaction
Input: RGB
Summary: Jointly learns 2D visual and 3D physical cues, then aggregates multiple diffusion-generated pose candidates to balance visual consistency and physical plausibility in hand–object results.
1 papers
Introduces security issues such as backdoor attacks into 6DoF pose estimation to examine failure boundaries under attack.
Jihui Guo · Zongmin Zhang · Zhen Sun · Yuhao Yang · Jinlin Wu · Fu Zhang · Xinlei He
Setting: Trigger-based attacks against existing 6DoF object-pose models
Input: Object-pose model inputs
Summary: Studies triggers that preserve normal 6DoF pose-model behavior on clean samples but produce attacker-specified incorrect rotations and translations on triggered samples, and analyzes existing defenses.
Scope & Sources
Inclusion: official AAAI-26 papers whose titles or abstracts directly address object 6D pose, category-level pose, articulated-object pose tracking, hand–object pose, or 6DoF pose-model security.
Excluded: human pose, relative camera / scene pose, 3D reconstruction alone, articulated-object modeling alone, Demo / Student Abstract entries, workshops, and general perception papers without direct object-pose outputs.
Primary sources: Official AAAI-26 proceedings archive, AAAI Proceedings and each paper's official page. Summaries are editorial explanations and do not replace the original papers.
