Research map / 3D vision / Human pose estimation
Any6D: Model-free 6D Pose Estimation of Novel Objects
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.
Any6D estimates relative 6D pose and metric size of novel objects from a single RGB-D anchor image by reconstructing a shape with InstantMesh and jointly aligning size and pose, then render-and-compare with FoundationPose-style refinement.
Problem
Model-free 6D pose and size estimation of novel objects from only a single RGB-D reference image, without CAD models or multi-view references, robust to occlusion, non-overlapping views and lighting changes.
Approach
Reconstruct a complete normalized object shape with an image-to-3D model, then jointly estimate metric size and pose via coarse OBB-based alignment and size-sampled render-and-compare refinement (adapted from FoundationPose), enabling full-to-partial matching instead of partial-to-partial point cloud matching.
Key results
- On HO3D, Any6D reaches mean ADD-S 98.7%, ADD 40.4%, AR 38.3%, versus Gedi's 71.9/9.7/7.4 and Oryon's 23.0/0.0/1.0 (ADD-S/ADD/AR).
- On REAL275, Any6D reports ADD(-S) 53.5% vs Oryon's 34.9%, and MSPD 65.3% vs 56.7%, though VSD is slightly lower (31.1 vs 32.1).
- On LM-O, Any6D with CNOS and Wonder3D gets AR 28.6 vs GigaPose's 17.5 under the same segmentation and image-to-3D inputs.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Any6D instead of Gedi | Model-free pose estimation · HO3D MEAN | ADD: 9.7 -> 40.4 | Helped |
| Use Any6D instead of Gedi | Model-free pose estimation · YCBInEOTA MEAN | ADD: 7.7 -> 45.6 | Helped |
| Use Any6D instead of Oryon | Model-free pose estimation · Toyota-Light | ADD(-S): 22.9 -> 32.2 | Helped |
| Use Any6D instead of Oryon | Model-free pose estimation · REAL275 | VSD: 32.1 -> 31.1 | Hurt |
| Remove all object alignment steps (ablation) | Any6D · HO3D | ADD: 40.4 -> 0.0 | Hurt |
| Add coarse size estimation (ablation) | Any6D without alignment · HO3D | ADD: 0.0 -> 26.6 | Helped |
| Add size and pose refinement (ablation) | Any6D with coarse size estimation and axis alignment · HO3D | ADD: 26.6 -> 40.4 | Helped |
| Add axis alignment (ablation) | Any6D with coarse size estimation and refinement · HO3D | ADD-S: 92.3 -> 98.7 | Helped |
Surprising findings
- Oryon's ADD scores are near zero on HO3D and YCBInEOAT, showing partial-matching methods fail under low overlap.
- Coarse size estimation alone raises ADD on HO3D from 0.0 to 25.5 in ablation, and the pose refinement adds further gains.
Limitations
- Performance depends on the accuracy of the initial image-to-3D shape; no shape updating is performed.
- Uses ground-truth segmentation masks for most datasets during evaluation.
- On REAL275 VSD is slightly below Oryon, and on some YCBInEOAT objects ADD-S is lower than Gedi.
Experimental setup
Datasets: DexYCB HO3D YCBInEOTA Toyota-Light REAL275 LM-O
Benchmarks: HO3D YCBInEOTA Toyota-Light REAL275 LM-O
Metrics: ADD ADD-S AR MSSD MSPD VSD AUC Chamfer Distance
Baselines: Oryon LoFTR Gedi ObjectMatch SIFT GigaPose
Builds on: FoundationPose Oryon InstantMesh GigaPose
Closest papers by meaning
- One2Any: One-Reference 6D Pose Estimation for Any Object (CVPR 2025 · similarity 0.91)
- SAM-6D: Segment Anything Model Meets Zero-Shot 6D Object Pose Estimation (CVPR 2024 · similarity 0.89)
- SingRef6D: Monocular Novel Object Pose Estimation with a Single RGB Reference (NeurIPS 2025 · similarity 0.89)
- Pos3R: 6D Pose Estimation for Unseen Objects Made Easy (CVPR 2025 · similarity 0.89)
- Open-Vocabulary Object 6D Pose Estimation (CVPR 2024 · similarity 0.89)
- iG-6DoF: Model-free 6DoF Pose Estimation for Unseen Object via Iterative 3D Gaussian Splatting (CVPR 2025 · similarity 0.89)
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.