Research map / Point clouds / Shape and surface completion
Occlusion-aware Non-Rigid Point Cloud Registration via Unsupervised Neural Deformation Correntropy
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 OAR, an unsupervised test-time non-rigid point cloud registration method combining a SIREN-style coordinate MLP with a maximum correntropy criterion and locally linear reconstruction regularization, targeting occluded inputs.
Problem
Non-rigid point cloud registration fails under occlusion, where Chamfer-distance-based objectives cause collapsed, torn, or physically implausible deformations when the source is complete and the target is partial.
Approach
Replace holistic Chamfer distance with a Gaussian-kernel correntropy (MCC) local similarity so occluded points contribute less, and regularize deformations with a closed-form locally linear reconstruction (LLR) prior instead of AIAP, all optimized unsupervised per shape pair with an implicit neural deformation field.
Key results
- On the occluded Open-CAS liver dataset, OAR reports EPE 8.662 / AccS 29.228 / AccR 96.813 on Liver 1 and zero outlier ratio across all three livers, beating all compared baselines.
- On 4DMatch with Lepard features, Lepard+OAR reaches AccS 59.32 and AccR 74.33 (best AccR, second AccS behind Lepard+NDP's 61.30 AccS).
- On occluded TOSCA cat/horse/dog, OAR achieves the best AccR on all three shapes (e.g., 89.779 on dog vs. 80.814 for DPF).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the LLR constraint to MCC (ablation) | OAR · Occluded lion shapes with progressively increasing occlusion | registration quality | Helped |
| Replace LLR with AIAP regularization (ablation) | MCC · Occluded shapes with different levels of occlusion | registration quality | Hurt |
| Add LLR to ReLU+MCC (ablation) | ReLU+MCC · Open-CAS liver 1 (occluded) | AccS: 28.446 -> 34.144 | Helped |
| Add LLR to ReLU+MCC (ablation) | ReLU+MCC · Open-CAS liver 2 (occluded) | AccR: 69.975 -> 89.956 | Helped |
| Replace CD with MCC (ablation) | ReLU deformation network · Open-CAS liver 1 (occluded) | AccS: 9.106 -> 28.446 | Helped |
| Replace CD with MCC (ablation) | ReLU deformation network · Open-CAS liver 2 (occluded) | AccR: 23.321 -> 69.975 | Helped |
| Replace ReLU with SIREN (ablation) | CD-based deformation network · Complete Open-CAS liver 1→liver 2 | EPE: 4.685 -> 3.912 | Helped |
| Increase the number of LLR neighbors k (ablation) | OAR · Occluded Open-CAS liver 1 | AccS | Helped |
| Vary the correntropy kernel bandwidth σ² (ablation) | OAR · Occluded Open-CAS liver 1 | registration performance | No clear effect |
| Use MCC instead of CD (ablation) | OAR · Occluded shapes with progressively increasing occlusion | AccR | Helped |
| Use MCC instead of HCD (ablation) | OAR on occluded Open-CAS livers · Open-CAS liver 1, liver 2, and liver 3 (occluded) | AccR | Helped |
| Use SIREN instead of ReLU (ablation) | CD-based deformation network · Complete Open-CAS liver shapes | EPE | No clear effect |
Surprising findings
- The paper reports that the regularization in occluded regions comes mainly from LLR rather than the network's activation function, since ReLU+CD and ReLU+MCC still show large errors.
Limitations
- Performance degrades when the occluded region is very large (failure cases with both arms and body occluded).
- LLR requires a matrix inverse per point, giving relatively high computational cost for large k.
- The problem is severely ill-posed, so deformation solutions are not unique under heavy occlusion.
Experimental setup
Datasets: Open-CAS silico liver dataset TOSCA (cat, horse, dog) 4DMatch 4DLoMatch DeformingThings4D
Benchmarks: Open-CAS liver occlusion TOSCA occlusion 4DMatch 4DLoMatch
Metrics: EPE AccS AccR Outlier ratio Time (s)
Baselines: GBCPD AMM_NRR NSFP NDP DPF NICP SyNoRiM Lepard+NDP HCD
Builds on: NSFP NDP DPF Lepard SIREN
Closest papers by meaning
- RGGT: A Generative-Prior-Guided Transformer for Unified Rigid and Non-Rigid Point Cloud Registration (ICML 2026 · similarity 0.90)
- Mining and Transferring Feature-Geometry Coherence for Unsupervised Point Cloud Registration (NeurIPS 2024 · similarity 0.87)
- NeuroGauss4D-PCI: 4D Neural Fields and Gaussian Deformation Fields for Point Cloud Interpolation (NeurIPS 2024 · similarity 0.87)
- DV-Matcher: Deformation-based Non-rigid Point Cloud Matching Guided by Pre-trained Visual Features (CVPR 2025 · similarity 0.87)
- Digging into Intrinsic Contextual Information for High-fidelity 3D Point Cloud Completion (AAAI 2025 · similarity 0.87)
- Rectified Point Flow: Generic Point Cloud Pose Estimation (NeurIPS 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.