Research map / 3D vision / Human pose estimation
UA-Pose: Uncertainty-Aware 6D Object Pose Estimation and Online Object Completion with Partial References
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes UA-Pose, which models partial object references as a mesh with seen/unseen uncertainty labels, uses it for confidence-aware 6D pose estimation, and performs online object completion with uncertainty-aware frame sampling, including single-image input via image-to-3D generation.
Problem
Model-free 6D object pose estimation fails when only partial references (few RGBD views or a single RGB image) are available, since existing methods assume complete 3D models or densely posed views.
Approach
A hybrid representation (texture, geometry, per-vertex seen/unseen uncertainty from visibility in reference masks) enables seen-IoU confidence scoring of poses, frame filtering, and uncertainty-driven selection of frames to complete the object model online during testing.
Key results
- On YCBInEOAT with 2 reference images, UA-Pose reaches mean 89.99 vs FoundationPose 68.52.
- On HO3D with 2 reference images, UA-Pose reaches mean 91.51 vs FoundationPose 45.67.
- On YCB-Video with 2 references, UA-Pose reaches mean 92.8 vs FoundationPose 87.4.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use UA-Pose with online object completion and uncertainty-aware pose estimation instead of FoundationPose | 6D pose estimation · YCB-Video, 2 references | Mean ADD: 87.4 -> 92.8 | Helped |
| use UA-Pose with online object completion and uncertainty-aware pose estimation instead of FoundationPose | 6D pose estimation · YCBInEOAT, 2 references | Mean ADD: 68.52 -> 89.99 | Helped |
| use UA-Pose with online object completion and uncertainty-aware pose estimation instead of FoundationPose | 6D pose estimation · HO3D, 2 references | Mean ADD: 45.67 -> 91.51 | Helped |
| use UA-Pose with online object completion and uncertainty-aware pose estimation instead of FoundationPose | 6D pose estimation · YCB-Video, single RGB + Image-to-3D | Mean ADD: 88.9 -> 93.2 | Helped |
| remove online object completion (ablation) | UA-Pose · YCBInEOAT, 2 references | Mean ADD: 89.99 -> 68.52 | Hurt |
| complete the model whenever a new test image enters the memory pool rather than only below the seen IoU threshold (ablation) | UA-Pose · YCBInEOAT, 2 references | Mean ADD: 89.99 -> 85.83 | Hurt |
| remove image filtering for unreliable poses (ablation) | UA-Pose · YCBInEOAT, 2 references | Mean ADD: 89.99 -> 87.70 | Hurt |
| replace uncertainty-aware sampling with geodesic distance-based sampling (ablation) | UA-Pose · YCBInEOAT, 2 references | Mean ADD: 89.99 -> 87.88 | Hurt |
Surprising findings
- Directly using image-to-3D generated models with FoundationPose yields lower ADD/ADD-S despite low Chamfer Distance, because generated textures in unseen regions are poor.
- Generated models can have better geometry (lower CD) than refined models yet still underperform for pose estimation.
Limitations
- Relies on off-the-shelf segmentation for test-frame masks and on sequential test frames.
- Single-image case depends on quality of image-to-3D generated models, whose unseen textures can be poor.
- Evaluated only on three datasets with RGBD inputs and a subset for the single-RGB setting.
Experimental setup
Datasets: YCB-Video YCBInEOAT HO3D
Benchmarks: YCB-Video YCBInEOAT HO3D
Metrics: ADD AUC ADD-S AUC Chamfer Distance mean AUC
Baselines: FoundationPose LoFTR FS6D-DPM BundleTrack BundleSDF
Builds on: FoundationPose BundleSDF neural SDF (NeuS-style volume rendering)
Closest papers by meaning
- 3D-Aware Hypothesis & Verification for Generalizable Relative Object Pose Estimation (ICLR 2024 · similarity 0.89)
- UNOPose: Unseen Object Pose Estimation with an Unposed RGB-D Reference Image (CVPR 2025 · similarity 0.89)
- Any6D: Model-free 6D Pose Estimation of Novel Objects (CVPR 2025 · similarity 0.89)
- One2Any: One-Reference 6D Pose Estimation for Any Object (CVPR 2025 · similarity 0.89)
- Pos3R: 6D Pose Estimation for Unseen Objects Made Easy (CVPR 2025 · similarity 0.88)
- 6D Object Pose Tracking in Internet Videos for Robotic Manipulation (ICLR 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.