Research map / 3D vision / Human pose estimation
RefPose: Leveraging Reference Geometric Correspondences for Accurate 6D Pose Estimation of Unseen 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.
RefPose estimates unseen-object 6D poses from monocular RGB by building a coarse pose from optical-flow-selected templates, then refining it with a correlation-volume-guided attention network that predicts query geometry against a rendered reference.
Problem
6D pose estimation of unseen objects from monocular RGB images, where lack of object-specific priors and poor generalization of geometric correspondence networks hurt accuracy.
Approach
Uses a coarse pose from template selection and medoid-voted warping to render an aligned reference image and geometry, then estimates query geometry conditioned on this reference via a correlation-volume-guided attention U-Net, followed by render-and-compare iterative refinement.
Key results
- Full RefPose achieves mean AR 61.4 across seven BOP datasets after refinement, versus 60.5 for the best prior (GigaPose with GenFlow multi-hypothesis refinement).
- Coarse pose estimation reaches mean AR 38.1, above FoundPose's 37.2.
- Runtime of 3.9s per image for the full pipeline, with the paper reporting competitive speed.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use RefPose instead of the strongest listed baseline | BOP benchmark pose estimation · BOP benchmark (mean) | AR: 60.5 -> 61.4 | Mixed |
| Use 64 pre-rendered templates instead of 128 (ablation) | RefPose coarse pose estimation · BOP benchmark (coarse pose, mean) | AR: 38.1 -> 33.9 | Hurt |
| Use 256 pre-rendered templates instead of 128 (ablation) | RefPose coarse pose estimation · BOP benchmark (coarse pose, mean) | AR: 38.1 -> 38.3 | No clear effect |
| Select 1 template instead of 4 (ablation) | RefPose coarse pose estimation · BOP benchmark (coarse pose, mean) | AR: 38.1 -> 31.5 | Hurt |
| Select 2 templates instead of 4 (ablation) | RefPose coarse pose estimation · BOP benchmark (coarse pose, mean) | AR: 38.1 -> 35.2 | Hurt |
| Select 8 templates instead of 4 (ablation) | RefPose coarse pose estimation · BOP benchmark (coarse pose, mean) | AR: 38.1 -> 36.2 | Hurt |
| Remove optical-flow network features from the template classifier (ablation) | RefPose coarse pose estimation · BOP benchmark (coarse pose, mean) | AR: 35.8 -> 38.1 | Hurt |
| Replace medoid voting with simple averaging (ablation) | RefPose coarse pose estimation · BOP benchmark (coarse pose, mean) | AR: 36.0 -> 38.1 | Hurt |
| Remove positional encoding from geometric correspondence (ablation) | RefPose pose refinement · BOP benchmark (refined pose, mean) | AR: 59.1 -> 61.4 | Hurt |
| Replace convex upsampling with bilinear upsampling (ablation) | RefPose pose refinement · BOP benchmark (refined pose, mean) | AR: 60.0 -> 61.4 | Hurt |
| Replace correlation volume-guided attention with vanilla attention (ablation) | RefPose pose refinement · BOP benchmark (refined pose, mean) | AR: 58.9 -> 61.4 | Hurt |
Surprising findings
- RefPose does not achieve the best results on LM-O, T-LESS, and ITODD, despite having the best overall mean AR.
Limitations
- Slightly underperforms state-of-the-art on LM-O, T-LESS, and ITODD.
- Coarse estimation takes slightly longer than some baselines.
- Trained only on GSO due to computational and memory constraints, omitting ShapeNet.
Experimental setup
Datasets: Google Scanned Objects (GSO)
Benchmarks: YCB-V LM-O T-LESS TUD-L ICBIN ITODD HB
Metrics: Average Recall (AR) VSD MSSD MSPD
Baselines: MegaPose GenFlow GigaPose FoundPose OSOP ZS6D
Builds on: RAFT MegaPose CNOS SCFlow NeRF positional encoding
Closest papers by meaning
- UNOPose: Unseen Object Pose Estimation with an Unposed RGB-D Reference Image (CVPR 2025 · similarity 0.91)
- SingRef6D: Monocular Novel Object Pose Estimation with a Single RGB Reference (NeurIPS 2025 · similarity 0.90)
- GeoReF: Geometric Alignment Across Shape Variation for Category-level Object Pose Refinement (CVPR 2024 · similarity 0.89)
- Co-op: Correspondence-based Novel Object Pose Estimation (CVPR 2025 · similarity 0.89)
- One2Any: One-Reference 6D Pose Estimation for Any Object (CVPR 2025 · similarity 0.88)
- MatchU: Matching Unseen Objects for 6D Pose Estimation from RGB-D Images (CVPR 2024 · 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.