AAAI 2025 · Paper Collection
AAAI 2025 has relatively few direct object-pose papers, but highlights three shifts: universal features extend category-level methods beyond training categories; articulated and freely moving objects reduce dependence on RGB-D or fixed priors; hand–object interaction brings runtime, contact, and data coverage into pose tasks.
6
Papers meeting the inclusion criteria
5
Method paper
3
Category-level / articulated parts
1
Data-augmentation work
These 6 papers meet this site's inclusion criteria: official AAAI-25 papers whose titles or abstracts directly address object 6D pose, category-level / articulated pose, freely moving object pose, or joint hand–object 3D pose. HOGSA is retained as data work directly supporting hand–object pose learning.
Collection in One Sentence
The core change in 2025 is the redesign of pipeline priors: Universal Features Pose uses universal 2D / 3D features to avoid category-specific retraining, R²-Art incorporates joint states into RGB-only refinement, and Virtual Camera explicitly reduces the global optimization search space for freely moving objects.
QORT-Former and HOGSA place pose within real interaction systems: the former limits Transformer query / decoder budgets for real-time inference, while the latter expands two-hand–object pose and viewpoint distributions. Deployment requires accuracy, speed, and data coverage to be considered together.
Research Trends
Trends summarize task conditions and method designs across the included papers; individual papers may span multiple approaches.
01 · Generalization
Universal Features Pose establishes sparse correspondences with 2D features and dense alignment with 3D features. KeyPose adapts keypoints to object geometry rather than imposing fixed keypoints on every instance.
02 · Dynamic Objects
R²-Art uses joint-centric priors and cascaded rendering for single-RGB articulated pose. Virtual Camera jointly optimizes monocular-video shape and pose without fixed-scene, hand, or category-pose priors.
03 · Interaction Systems
QORT-Former assigns separate queries to both hands and objects and includes contact in a real-time Transformer. HOGSA expands two-hand–object pose / viewpoint coverage, addressing difficult interaction-data acquisition and heavy occlusion.
Data Statistics
Each paper belongs to one group by its main problem setting; card keywords retain cross-theme relationships.
Handles within-category variation through RGB-D, general-purpose features, and adaptive keypoints, extending pose estimation toward categories unseen during training.
Jointly addresses articulation states, object shapes, and continuous-motion pose from single RGB images or monocular video.
Combines two hands, objects, contact relationships, and real-time requirements in a unified 3D pose / interaction-understanding framework.
Expands viewpoint and pose distributions in hand–object interaction data through controllable 3DGS rendering and grasp-pose optimization.
Interpreting the statistics: Category-level and articulated-part estimation remain central to direct object pose. RGB-only methods, monocular video, real-time interaction, and 3DGS augmentation reduce constraints from depth, fixed priors, latency, and data scale. AAAI-25 emphasizes pose in dynamic, interactive environments beyond a single BOP ranking.
Recommended Reading
Recommendations reflect problem coverage and methodological representativeness, rather than official AAAI rankings or best-paper selections.
Integrates 2D / 3D foundation-style universal features into a complete zero-shot pipeline from sparse coarse pose to dense refinement, representing AAAI-25's most direct general-purpose category-level pose work.
Uses object-adaptive geometric anchors instead of fixed semantic keypoints, with graph aggregation and NOCS voxelization for pose fitting, restructuring the conventional category-level approach.
Connects RGB-only articulated pose, joint-centric prior deformation, and cascaded rendering in a closed loop, representing a shift from RGB-D / point-cloud dependence to monocular settings.
Uses virtual cameras to reduce the global sequence-optimization search space without scene, hand, or category-pose priors, illustrating the shift toward dynamic-object modeling.
Explicitly optimizes query and decoder counts in a Transformer shared across hands, objects, and interaction categories, treating real-time efficiency as a primary constraint.
Paper List
Each card opens the site's detail page with coarse and deep notes, original method figures, experimental analysis, and official external links.
2 papers
Handles within-category variation through RGB-D, general-purpose features, and adaptive keypoints, extending pose estimation toward categories unseen during training.
Wentian Qu · Chenyu Meng · Heng Li · Jian Cheng · Cuixia Ma · Hongan Wang · Xiao Zhou · Xiaoming Deng · Ping Tan
Setting: Category-level 6D pose extending to unseen categories without additional fine-tuning
Input: RGB-D
Summary: First establishes sparse semantic correspondences with universal 2D features for coarse pose, then iteratively corrects shape-induced pose ambiguity through dense alignment of universal 3D features. Evaluates unseen categories on REAL275 and Wild6D.
Sheng Yu · Di-Hua Zhai · Yuanqing Xia
Setting: Category-level 6D pose for unseen instances with substantial appearance variation
Input: RGB + point cloud
Summary: A Transformer predicts geometry-adaptive keypoints, aggregates structural information through keypoint-to-keypoint and keypoint-to-cloud graphs, then predicts canonical keypoints accurately through NOCS voxelization to recover 6D pose.
2 papers
Jointly addresses articulation states, object shapes, and continuous-motion pose from single RGB images or monocular video.
Li Zhang · Haonan Jiang · Yukang Huo · Yan Zhong · Jianan Wang · Xue Wang · Rujing Wang · Liu Liu
Setting: Category-level articulated-object part 6D poses from one RGB image
Input: RGB
Summary: Initializes through parallel camera-to-root translation and rotation regression, transforms / deforms part priors using predicted joint states, then projects 3D priors into 2D masks through cascaded rendering for refinement.
Haixin Shi · Yinlin Hu · Daniel Koguciuk · Juan-Ting Lin · Mathieu Salzmann · David Ferstl
Setting: Joint shape and continuous-pose optimization for freely interacting objects in monocular video
Input: Monocular RGB video
Summary: Progressively jointly optimizes implicit object shape and pose from monocular sequences with freely moving cameras and objects. A virtual camera system reduces the global search space, evaluated on HO3D and egocentric RGB sequences.
1 papers
Combines two hands, objects, contact relationships, and real-time requirements in a unified 3D pose / interaction-understanding framework.
Elkhan Ismayilzada · MD Khalequzzaman Chowdhury Sayem · Yihalem Yimolal Tiruneh · Mubarrat Tajoar Chowdhury · Muhammadjon Boboev · Seungryul Baek
Setting: Joint two-hand and object 3D pose and interaction understanding
Input: RGB video
Summary: Assigns separate query types to the left hand, right hand, and object. Hand–object contact and three-step image–query interaction improve representations, enabling real-time joint 3D pose with only 108 queries and 1 decoder.
1 papers
Expands viewpoint and pose distributions in hand–object interaction data through controllable 3DGS rendering and grasp-pose optimization.
Wentian Qu · Jiahe Li · Jian Cheng · Jian Shi · Chenyu Meng · Cuixia Ma · Hongan Wang · Xiaoming Deng · Yinda Zhang
Setting: Pose / viewpoint distribution augmentation for two-hand–object interaction understanding
Input: multi-view hand-object capture / synthetic render
Summary: Reconstructs hands and objects through mesh-based 3DGS, uses super-resolution to reduce rendering blur from multiresolution acquisition, and extends single-hand grasp-pose optimization to two-hand–object scenes for additional photorealistic poses and viewpoints.
Scope & Sources
Inclusion: official AAAI-25 Technical Track papers whose titles or abstracts directly address object 6D pose, category-level pose, articulated pose, freely moving object pose / shape, or joint hand–object 3D pose. HOGSA data augmentation is retained for directly supporting hand–object pose learning.
Excluded: human pose, relative camera / scene pose, object reconstruction without pose recovery, action recognition alone, papers mentioning pose only as downstream context, and Demo / Student Abstract / Workshop entries.
Primary sources: Official AAAI-25 conference archive, AAAI Proceedings Technical Track 6 and each paper's official page. Summaries are editorial explanations and do not replace the original papers.
