Research map / Point clouds / Registration and matching
Inlier Confidence Calibration for 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 inlier confidence calibration (ICC) for unsupervised partial-overlap point cloud registration, scoring correspondences via transformation-invariant geometric consistency between source and a reference copy, plus a faster row-only soft assignment theorem replacing Sinkhorn.
Problem
Inlier estimation in partially overlapping point cloud registration; coordinate-difference-based inlier confidence leads to inlier misinterpretation and degraded registration accuracy.
Approach
Score inlier confidence by checking consistency of transformation-invariant geometric structure (edge lengths and included angles) between each source point's neighborhood and its counterpart neighborhood in a reference point cloud copy, and use the calibrated confidence in weighted SVD; the geometric consistency also serves as a self-supervised loss.
Key results
- On ModelNet40 unseen objects, the method reaches MAE(R) 0.0012 and MIE(R) 0.0192, beating the prior unsupervised RIE (MAE(R) 0.0033) and supervised methods such as RPMNet (0.0051).
- On Augmented ICL-NUIM the method reports MAE(R) 0.0048 versus 0.0492 for RIE; on 7Scenes MAE(R) 0.0033 versus 0.0121 for RIE.
- The row-only soft assignment theorem is reported to be up to about 2x faster than Sinkhorn-1 and up to about 5x faster than Sinkhorn-5, with complexity O(N) versus O(G x N x M).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add transformation-invariant geometric constraints for inlier confidence calibration (ablation) | ICC · ModelNet40 | MAE(R): 1.0123 -> 0.0012 | Helped |
| add transformation-invariant geometric constraints for inlier confidence calibration (ablation) | ICC · 7Scenes | MAE(R): 2.2238 -> 0.0033 | Helped |
| remove the global consistency loss Lgc (ablation) | ICC · ModelNet40 | MAE(R): 0.0012 -> 0.3998 | Hurt |
| remove the inlier neighborhood consistency loss Lin (ablation) | ICC · ModelNet40 | MAE(R): 0.0012 -> 22.5199 | Hurt |
| remove the geometric structure consistency loss Lgs (ablation) | ICC · ModelNet40 | MAE(R): 0.0012 -> 0.0070 | Hurt |
| remove the spatial consistency loss Lsc (ablation) | ICC · ModelNet40 | MAE(R): 0.0012 -> 0.1125 | Hurt |
| replace Sinkhorn with the proposed soft assignment matrix optimization theorem | ICC · ModelNet40 | rotation error | Helped |
| replace Sinkhorn with the proposed soft assignment matrix optimization theorem | ICC · ModelNet40 | runtime: nearly 2×; nearly up to 5× | Helped |
Surprising findings
- Removing the geometric constraints and using coordinate-based confidence yields a much larger error (MAE(R) 1.0123 on ModelNet40 without constraints), supporting the claim that coordinate-based scoring misleads inlier estimation.
Limitations
- Gaussian-distribution-based theorem justification is heuristic; the bound on row-wise dominance is only argued through the three-sigma rule.
- Results are reported only on synthetic object datasets and two indoor scans; cross-modality (2D-3D) registration is left as future work.
Experimental setup
Datasets: ModelNet40 Augmented ICL-NUIM 7Scenes
Benchmarks: ModelNet40 Augmented ICL-NUIM 7Scenes
Metrics: MAE(R) MAE(t) MIE(R) MIE(t)
Baselines: ICP FGR FPFH+RANSAC IDAM RPMNet FMR REGTR GeoTransformer CEMNet RIE
Builds on: RIE RPM-Net Sinkhorn Dynamic Graph CNN GeoTransformer
Closest papers by meaning
- Reliable Inlier Evaluation for Unsupervised Point Cloud Registration (AAAI 2022 · similarity 0.94)
- A robust inlier identification algorithm for point cloud registration via $\mathbf{\ell_0}$-minimization (NeurIPS 2024 · similarity 0.94)
- PointDSC: Robust Point Cloud Registration Using Deep Spatial Consistency (CVPR 2021 · similarity 0.89)
- Registration is a Powerful Rotation-Invariance Learner for 3D Anomaly Detection (NeurIPS 2025 · similarity 0.85)
- Multi-scale Consistency for Robust 3D Registration via Hierarchical Sinkhorn Tree (NeurIPS 2024 · similarity 0.85)
- Cross-PCR: A Robust Cross-Source Point Cloud Registration Framework (AAAI 2025 · 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.