Research map / Point clouds / Registration and matching
Registration is a Powerful Rotation-Invariance Learner for 3D Anomaly Detection
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 Reg2Inv, which trains a registration network (RIConv++ + KPConv-FPN + Geometric Transformer) whose features are reused for rotation-invariant memory-bank 3D anomaly detection on Real3D-AD and Anomaly-ShapeNet.
Problem
Memory bank-based 3D point-cloud anomaly detection suffers from residual misalignment after FPFH+RANSAC registration and from feature encoders (e.g., PointMAE) that lack rotation invariance and local geometric sensitivity.
Approach
Treat point-cloud registration as the feature-learning objective: train a rotation-invariant extractor with registration losses (feature, point and overlap-aware circle losses), then use the same features and registration matrix for memory-bank anomaly scoring.
Key results
- On Real3D-AD, average O-AUROC 78.0% and P-AUROC 87.8%, outperforming the second-best method by 1.5% (O-AUROC) and 4.2% (P-AUROC).
- On Anomaly-ShapeNet, achieves best O-AUROC 86.1% (2.2% above second-best) and P-AUROC 88.2%, second to PO3AD's 89.8%.
- Ablation: removing the overlap-aware circle loss drops Real3D-AD O-AUROC to 58.1% and P-AUROC to 68.1%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Reg2Inv instead of ISMP | 3D anomaly detection · Real3D-AD | O-AUROC: 75.7 -> 78.0 | Helped |
| Use Reg2Inv instead of ISMP | 3D anomaly detection · Real3D-AD | P-AUROC: 83.6 -> 87.8 | Helped |
| Use Reg2Inv instead of PO3AD | 3D anomaly detection · Anomaly-ShapeNet | O-AUROC: 83.9 -> 86.1 | Helped |
| Use Reg2Inv instead of PO3AD | 3D anomaly detection · Anomaly-ShapeNet | P-AUROC: 89.8 -> 88.2 | Hurt |
| Use proposed registration instead of FPFH+RANSAC (ablation) | Reg2Inv · Real3D-AD | P-AUROC: 72.6 -> 87.8 | Helped |
| Add L_oc to the training objective (ablation) | Reg2Inv · Real3D-AD | O-AUROC: 58.1 -> 78.0 | Helped |
| Add L_oc to the training objective (ablation) | Reg2Inv · Real3D-AD | P-AUROC: 68.1 -> 87.8 | Helped |
| Add L_p to the training objective (ablation) | Reg2Inv · Real3D-AD | O-AUROC: 76.9 -> 78.0 | Helped |
| Add L_f to the training objective (ablation) | Reg2Inv · Real3D-AD | P-AUROC: 86.0 -> 87.8 | Helped |
| Combine coordinates and local features instead of coordinates alone (ablation) | Reg2Inv memory bank · Real3D-AD | P-AUROC: 66.2% -> 87.8% | Helped |
| Combine coordinates and local features instead of coordinates alone (ablation) | Reg2Inv memory bank · Anomaly-ShapeNet | P-AUROC: 67.2% -> 88.2% | Helped |
| Use normalization and filtering rather than neither (ablation) | Reg2Inv inference · Real3D-AD | O-AUROC: 60.1 -> 78.0 | Helped |
Surprising findings
- Reported Real3D-AD P-AUROC for the method is far below PO3AD's on several categories, and the method's Real3D-AD O-AUROC on Fish and Seahorse is lower than baselines.
- Rotation test shows nearly identical O-AUROC/P-AUROC (76.2-76.6 / 83.6-84.3) across 0/90/180/270 degree rotations.
Limitations
- Requires training a separate model per class, increasing computational and storage costs (per the paper's limitations, though appendix says category-agnostic training).
- Limited generalization in less discriminative classes for the rotation-invariant extractor.
- Registration is not always accurate for anomaly detection.
Experimental setup
Datasets: Anomaly-ShapeNet Real3D-AD
Benchmarks: Real3D-AD Anomaly-ShapeNet
Metrics: O-AUROC P-AUROC AITPS
Baselines: IMRNet R3D-AD PO3AD BTF M3DM PatchCore CPMF Reg3D-AD Group3AD ISMP
Builds on: RIConv++ Geometric Transformer PatchCore Reg3D-AD Sinkhorn
Closest papers by meaning
- Towards Scalable 3D Anomaly Detection and Localization: A Benchmark via 3D Anomaly Synthesis and A Self-Supervised Learning Network (CVPR 2024 · similarity 0.90)
- Towards Generalizable 3D Anomaly Detection via Relational Inconsistency Modeling (NeurIPS 2026 · similarity 0.88)
- PO3AD: Predicting Point Offsets toward Better 3D Point Cloud Anomaly Detection (CVPR 2025 · similarity 0.88)
- A robust inlier identification algorithm for point cloud registration via $\mathbf{\ell_0}$-minimization (NeurIPS 2024 · similarity 0.87)
- Looking 3D: Anomaly Detection with 2D-3D Alignment (CVPR 2024 · similarity 0.87)
- Mining and Transferring Feature-Geometry Coherence for Unsupervised Point Cloud Registration (NeurIPS 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.