Research map / 3D vision / Human pose estimation
Pos3R: 6D Pose Estimation for Unseen Objects Made Easy
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.
Pos3R estimates 6D poses of unseen objects from one RGB image in a training-free way by matching CAD-rendered templates to test crops with the 3D foundation model MASt3R, then solving PnP-RANSAC.
Problem
Model-based 6D pose estimation for unseen objects typically needs task-specific training or relies on 2D foundation features that handle out-of-plane rotation poorly, and template selection is a bottleneck.
Approach
Use MASt3R's 3D-consistent dense correspondences between test crops and only 40 templates (8 cube-vertex views times 5 in-plane rotations), selecting the template by summed feature similarity of reciprocal matches, then fitting pose with PnP-RANSAC on 2D-3D correspondences.
Key results
- Pos3R achieves a mean BOP AR of 39.5 across seven datasets for coarse estimation, versus 37.2 for FoundPose and 32.3 for the no-axial-rotation baseline, at 1.4 s per image.
- Replacing MASt3R with DINOv2 (layer 11) drops the mean-style accuracy sharply, e.g., LM-O from 32.3 to 18.5 and YCB-V from 53.9 to 22.5.
- With MegaPose refinement on five hypotheses, Pos3R reaches a mean AR of 57.3, close to FoundPose (56.8) and GigaPose (57.9).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MASt3R rather than DINOv2 features (ablation) | Pos3R · BOP datasets | AR: Pos3R 32.3, 47.3, 53.9, 31.5, 33.1, 25.1, 53.7; DINOv2 (L11) 18.5, 25.0, 22.5, 14.2, 15.6, 14.1, 32.1 | Helped |
| Add axial rotation augmentation (ablation) | Pos3R · BOP datasets | AR: Pos3R 32.3, 47.3, 53.9, 31.5, 33.1, 25.1, 53.7; w/o Ax. Rot. 30.5, 43.2, 49.3, 23.6, 21.4, 15.4, 42.5 | Helped |
| Use cube-vertex base templates rather than face-center templates (ablation) | Pos3R · BOP datasets | AR: Pos3R 32.3, 47.3, 53.9, 31.5, 33.1, 25.1, 53.7; Face Center 31.5, 45.5, 53.3, 28.0, 29.4, 22.4, 52.5 | Helped |
| Replace MASt3R with RoMa (ablation) | Pos3R · BOP datasets | AR: Pos3R 32.3, 47.3, 53.9, 31.5, 33.1, 25.1, 53.7; RoMa 20.8, 25.0, 32.0, 16.2, 17.0, 15.5, 25.1 | Hurt |
| Replace MASt3R with DINOv2 (L9) (ablation) | Pos3R · BOP datasets | AR: Pos3R 32.3, 47.3, 53.9, 31.5, 33.1, 25.1, 53.7; DINOv2 (L9) 20.5, 26.0, 24.0, 15.6, 16.6, 14.8, 32.9 | Hurt |
| Select templates by correspondence similarity rather than confidence (ablation) | Pos3R · BOP datasets | AR | Helped |
| Select templates by correspondence similarity rather than inlier count (ablation) | Pos3R · BOP datasets | AR | Helped |
| Use Pos3R rather than FoundPose for coarse pose estimation | training-free unseen-object pose estimation · BOP seven-dataset mean | AR: 37.2 -> 39.5 | Helped |
| Use Pos3R rather than FoundPose for coarse pose estimation | training-free unseen-object pose estimation · LM-O | AR: 39.6 -> 32.3 | Hurt |
| Apply MegaPose refinement to Pos3R | Pos3R · BOP seven-dataset mean | AR: 39.5 -> 57.3 | Helped |
| Use Wonder3D-predicted 3D models instead of GT 3D models | Pos3R · LM-O | AR: 32.3 -> 19.5 | Hurt |
Surprising findings
- MASt3R-based matching outperforms RoMa, a dedicated dense matcher, on the same pipeline (e.g., YCB-V 53.9 vs 32.0).
Limitations
- Performance drops under heavy occlusion, notably on LM-O (32.3 AR), where FoundPose is stronger.
- Not inherently designed for refinement; needs MegaPose refiner (trained with pose supervision) to reach refined-level accuracy.
Experimental setup
Datasets: LM-O T-LESS TUD-L IC-BIN ITODD HomebrewedDB YCB-Video
Benchmarks: BOP LM-O BOP T-LESS BOP TUD-L BOP IC-BIN BOP ITODD BOP HomebrewedDB BOP YCB-V
Metrics: Average Recall (AR) VSD AR MSSD AR MSPD AR runtime (s)
Baselines: GigaPose GenFlow MegaPose OSOP ZS6D FoundPose DINOv2-based matching RoMa
Builds on: MASt3R CNOS MegaPose FoundPose GigaPose
Closest papers by meaning
- SingRef6D: Monocular Novel Object Pose Estimation with a Single RGB Reference (NeurIPS 2025 · similarity 0.89)
- Any6D: Model-free 6D Pose Estimation of Novel Objects (CVPR 2025 · similarity 0.89)
- iG-6DoF: Model-free 6DoF Pose Estimation for Unseen Object via Iterative 3D Gaussian Splatting (CVPR 2025 · similarity 0.89)
- One2Any: One-Reference 6D Pose Estimation for Any Object (CVPR 2025 · similarity 0.88)
- UA-Pose: Uncertainty-Aware 6D Object Pose Estimation and Online Object Completion with Partial References (CVPR 2025 · similarity 0.88)
- Co-op: Correspondence-based Novel Object Pose Estimation (CVPR 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.