Research map / 3D vision / Human pose estimation
DVMNet: Computing Relative Pose for Unseen Objects Beyond Hypotheses
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 DVMNet, a hypothesis-free deep voxel matching network that lifts reference and query RGB images to voxels and computes relative object rotation in one pass via a weighted closest voxel least-squares solver.
Problem
Relative pose estimation of unseen objects from a single reference image and a query image; existing hypothesis-based methods require many discrete pose hypotheses that are costly to score and produce unnaturally large errors.
Approach
Replace pose hypothesis scoring with differentiable voxel-to-voxel matching: an autoencoder with cross-view attention lifts 2D ViT features to 3D voxels, soft-aligned voxels feed a weighted Procrustes/SVD least-squares solver (WCV) weighted by 2D masks and 3D objectness maps, trained end-to-end on the pose loss.
Key results
- On CO3D, DVMNet reaches 19.95 degree mean angular error versus 28.44 for 3DAHV and 32.82 for RelPose++, an improvement of at least 8.49 degrees.
- On the GROP Objaverse split, DVMNet achieves 20.19 degree mean error versus 28.11 for 3DAHV, and on LINEMOD 36.82 versus 41.55 for 3DAHV.
- DVMNet processes an image pair in 23 ms on an A100, versus 29 ms for RelPose++ and 35 ms for 3DAHV.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DVMNet instead of 3DAHV | Relative object pose estimation · CO3D | angular error: 28.44 -> 19.95 | Helped |
| Use DVMNet instead of 3DAHV | Relative object pose estimation · LINEMOD | angular error: 41.55 -> 36.82 | Helped |
| Use DVMNet instead of 3DAHV | Relative object pose estimation · Objaverse | angular error: 28.11 -> 20.19 | Helped |
| Use WCV instead of pose regression (ablation) | DVMNet · CO3D | angular error: 31.78 -> 19.95 | Helped |
| Add voxel-pair weights to the closest-voxel algorithm (ablation) | DVMNet · CO3D | angular error: 21.64 -> 19.95 | Helped |
| Add voxel objectness to WCV weights (ablation) | DVMNet · CO3D | angular error: 20.92 -> 19.95 | Helped |
| Add the 2D object mask to WCV weights (ablation) | DVMNet · CO3D | angular error: 20.07 -> 19.95 | Helped |
Surprising findings
- Replacing the WCV algorithm with a direct pose regression head increases CO3D mean angular error from 19.95 to over 30 degrees, showing regression alone is much worse than the least-squares formulation.
Limitations
- Relies on ground-truth object bounding boxes to crop objects, so rotation-only estimation is evaluated.
- Translation is not addressed; authors defer it to zero-shot object detectors.
- Hypothesis-based baselines in MACs comparison were capped (e.g., 3DAHV at 100,000 samples) due to compute limits.
Experimental setup
Datasets: CO3D Objaverse LINEMOD
Benchmarks: CO3D LINEMOD Objaverse
Metrics: mean angular error MACs (G) runtime (ms)
Baselines: SuperGlue LoFTR ZSP Regress RelPose RelPose++ 3DAHV
Builds on: RelPose++ 3DAHV LoFTR SuperGlue
Closest papers by meaning
- 3D-Aware Hypothesis & Verification for Generalizable Relative Object Pose Estimation (ICLR 2024 · similarity 0.88)
- UNOPose: Unseen Object Pose Estimation with an Unposed RGB-D Reference Image (CVPR 2025 · similarity 0.86)
- Relative Pose Estimation through Affine Corrections of Monocular Depth Priors (CVPR 2025 · similarity 0.86)
- From Correspondences to Pose: Non-minimal Certifiably Optimal Relative Pose without Disambiguation (CVPR 2024 · similarity 0.86)
- FAR: Flexible Accurate and Robust 6DoF Relative Camera Pose Estimation (CVPR 2024 · similarity 0.86)
- NOPE: Novel Object Pose Estimation from a Single Image (CVPR 2024 · similarity 0.86)
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.