Research map / 3D vision / Human pose estimation
HiPose: Hierarchical Binary Surface Encoding and Correspondence Pruning for RGB-D 6DoF Object Pose Estimation
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 HiPose, a RANSAC-free, render-free RGB-D 6DoF object pose method that predicts hierarchical binary surface codes per point and iteratively prunes correspondences using coarse-to-fine point-to-surface matching.
Problem
Estimating 6DoF object pose from a single RGB-D image accurately and quickly, without costly rendering-based or ICP refinement, under occlusion and symmetric or textureless objects.
Approach
Predict a d-bit hierarchical binary surface code per 3D point, use the first m bits for a coarse pose via Kabsch on sub-surface centroids, then iteratively partition the surface with the remaining bits and prune outlier correspondences using distances to the transformed sub-surface, with per-bit confidence from the network output to choose the initial bit; no RANSAC or rendering is needed.
Key results
- On LM-O, HiPose reaches 89.6% mean ADD(-S) recall vs 77.7% for the best RGB-D baseline DFTr (per the reported table).
- On YCB-V, HiPose achieves 95.5 AUC of ADD(-S) and 97.6 AUC of ADD-S, outperforming DFTr (94.4).
- On the BOP benchmark average over LM-O, YCB-V, and T-LESS, HiPose scores 84.6 with 0.16 s per-object pose time, without refinement.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use hierarchical correspondence pruning instead of direct Kabsch pose solving (ablation) | HiPose · LM-O | ADD(-S) recall (%): 86.65 -> 89.62 | Helped |
| Replace hierarchical pruning with RANSAC+Kabsch (ablation) | HiPose · LM-O | ADD(-S) recall (%): 89.62 -> 89.12 | Hurt |
| Set trust-bit threshold to 0.08, 0.06, or 0.04 (ablation) | HiPose · LM-O | ADD(-S) recall (%): 89.62 -> 89.55; 89.62 -> 89.61; 89.62 -> 89.59 | No clear effect |
| Use the mean instead of the median for the inlier threshold (ablation) | HiPose · LM-O | ADD(-S) recall (%): 89.62 -> 89.47 | Hurt |
| Replace ConvNext-B with ResNet34 (ablation) | HiPose · LM-O | ADD(-S) recall (%): 89.62 -> 88.64 | Hurt |
| Use ZebraPose correspondences with RANSAC+Kabsch (ablation) | HiPose · LM-O | ADD(-S) recall (%): 89.62 -> 87.0 | Hurt |
| Concatenate RGB and depth as input and use RANSAC+Kabsch (ablation) | HiPose · LM-O | ADD(-S) recall (%): 89.62 -> 84.5 | Hurt |
| Use HiPose instead of the strongest listed RGB-D baseline | HiPose · LM-O | ADD(-S) recall (%): 77.7 -> 89.6 | Helped |
| Use HiPose instead of RCVPose | HiPose · YCB-V | AUC of ADD(-S) (%): 95.2 -> 95.5 | Helped |
| Use HiPose instead of DFTr | HiPose · YCB-V | AUC of ADD-S (%): 96.7 -> 97.6 | Helped |
| Use HiPose instead of the strongest listed BOP baseline | HiPose · T-LESS | BOP score: 85.2 -> 83.3 | Hurt |
| Use hierarchical correspondence pruning across iterations | HiPose · LM-O | Outlier-removal precision (%): 66.1 -> 76.1 | Helped |
Surprising findings
- RANSAC+Kabsch with tuned parameters only reaches 89.12 on LM-O, while the RANSAC-free hierarchical pruning reaches 89.62 and is stable across random seeds.
Limitations
- Trained only on synthetic PBR data in the BOP setting; T-LESS result is slightly below some refinement-based methods (83.3 vs 85.2 for GDRNPP).
- Relies on a 2D detector (CDPN/FCOS or YOLOX) for RoI cropping, adding detection time.
Experimental setup
Datasets: BOP PBR synthetic images LM-O YCB-V T-LESS
Benchmarks: LM-O YCB-V T-LESS
Metrics: ADD(-S) recall AUC of ADD-S AUC of ADD(-S) BOP score runtime
Baselines: ZebraPose FFB6D RCVPose DFTr GDR-Net PVN3D SO-Pose SurfEmb RADet+PFA GDRNPP PR-GCN
Builds on: ZebraPose FFB6D Kabsch algorithm ConvNeXt CIR
Closest papers by meaning
- SingRef6D: Monocular Novel Object Pose Estimation with a Single RGB Reference (NeurIPS 2025 · similarity 0.88)
- Co-op: Correspondence-based Novel Object Pose Estimation (CVPR 2025 · similarity 0.87)
- RefPose: Leveraging Reference Geometric Correspondences for Accurate 6D Pose Estimation of Unseen Objects (CVPR 2025 · similarity 0.87)
- MatchU: Matching Unseen Objects for 6D Pose Estimation from RGB-D Images (CVPR 2024 · similarity 0.87)
- GigaPose: Fast and Robust Novel Object Pose Estimation via One Correspondence (CVPR 2024 · similarity 0.86)
- Pos3R: 6D Pose Estimation for Unseen Objects Made Easy (CVPR 2025 · 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.