Research map / Point clouds / Registration and matching
HeMoRa: Unsupervised Heuristic Consensus Sampling for Robust Point Cloud Registration
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 HeMoRa, an unsupervised framework that trains a heuristic generator (HeGen) for correspondence sampling via a policy-gradient-style MoRa loss with max-pooling reward propagation, improving consensus-based point cloud registration.
Problem
Learning heuristic sampling probabilities for consensus-set sampling in correspondence-based point cloud registration without ground-truth transformations, where existing learned methods require supervised labels and classical methods need manual threshold tuning.
Approach
Train a sampling-probability and feature generator with a policy-gradient-like loss in which rewards (inlier ratio of sampled hypotheses) are propagated to all correspondences consistent with each hypothesis and aggregated by max-pooling instead of mean, plus a second-order pairwise term over correspondence pairs.
Key results
- On 3DMatch with FCGF, Ours+SC2 reaches 93.65% registration recall, above supervised VBReg (93.10%) and SC2-PCR (93.28%).
- On 3DLoMatch with FCGF, Ours+SC2 achieves 58.5% RR at 5000 keypoints, a 1.1% gain over SC2-PCR.
- Max-pooling aggregation raises Ours+SM RR on 3DMatch FPFH from 78.13% to 80.16% over mean pooling (ablation).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace sampled-correspondence indicator Iq-V1 with consistency-aware Iq-V2 reward propagation (ablation) | HeGen with mean aggregation; Ours+SM on 3DMatch (FPFH) · 3DMatch (FPFH) | RR (%): 77.33% -> 78.13% | Helped |
| Replace sampled-correspondence indicator Iq-V1 with consistency-aware Iq-V2 reward propagation (ablation) | HeGen with mean aggregation; Ours+PSAC on 3DMatch (FPFH) · 3DMatch (FPFH) | RR (%): 68.33% -> 68.88% | Helped |
| Replace mean pooling with max pooling (ablation) | HeGen with Iq-V2; Ours+SM on 3DMatch (FPFH) · 3DMatch (FPFH) | RR (%): 78.13% -> 80.16% | Helped |
| Replace mean pooling with max pooling (ablation) | HeGen with Iq-V2; Ours+PSAC on 3DMatch (FPFH) · 3DMatch (FPFH) | RR (%): +3.27% | Helped |
| Add second-order MoRa loss L(2) (ablation) | Ours+SM on 3DMatch (FPFH) · 3DMatch (FPFH) | RR (%): +1.23% | Helped |
| Add second-order MoRa loss L(2) (ablation) | Ours+PSAC and Ours+SC2 on 3DMatch (FPFH) · 3DMatch (FPFH) | RR (%) | No clear effect |
Surprising findings
- Ours+PSAC underperforms RANSAC-based baselines on some settings, e.g. KITTI-20m FPFH RR 20.32% versus 65% for MAC.
- The unsupervised method outperforms supervised PointDSC and VBReg in cross-descriptor generalization on 3DMatch.
Limitations
- Ours+PSAC performs poorly on some low-overlap KITTI settings, with large drops relative to baselines.
- Results depend on inlier threshold and Q settings chosen per descriptor (e.g. Q=1024 vs 8192).
Experimental setup
Datasets: 3DMatch 3DLoMatch KITTI-10m
Benchmarks: 3DMatch 3DLoMatch KITTI-10m KITTI-20m KITTI-30m
Metrics: registration recall rotation error translation error
Baselines: TEASER++ GC-RANSAC RANSAC CG-SAC SC2-PCR MAC 3DRegNet DGR DHVR PointDSC VBReg
Builds on: RANSAC PointDSC SC2-PCR policy gradient (REINFORCE)
Closest papers by meaning
- RPM-Net: Robust Point Matching Using Learned Features (CVPR 2020 · similarity 0.87)
- MHopReg: Efficient Hierarchical Multi-Hop Graph Search for Point Cloud Registration (CVPR 2026 · similarity 0.87)
- SDRSAC: Semidefinite-Based Randomized Approach for Robust Point Cloud Registration Without Correspondences (CVPR 2019 · similarity 0.87)
- Cross-PCR: A Robust Cross-Source Point Cloud Registration Framework (AAAI 2025 · similarity 0.86)
- PointMC: Multi-instance Point Cloud Registration based on Maximal Cliques (ICML 2024 · similarity 0.85)
- FreeReg: Image-to-Point Cloud Registration Leveraging Pretrained Diffusion Models and Monocular Depth Estimators (ICLR 2024 · similarity 0.85)
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.