Research map / Point clouds / Point cloud classification
Towards Scalable 3D Anomaly Detection and Localization: A Benchmark via 3D Anomaly Synthesis and A Self-Supervised Learning Network
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.
Introduces Anomaly-ShapeNet, a synthetic 3D point-cloud anomaly detection benchmark built from ShapeNetCoreV2 with Blender-carved defects, and IMRNet, a self-supervised iterative point-patch mask reconstruction method for 3D anomaly detection and localization.
Problem
Point-cloud 3D anomaly detection is limited by the scarcity of diverse real 3D anomaly data and by existing methods that rely on pretrained features from other domains, causing feature domain misalignment.
Approach
Synthesize 3D anomalies on ShapeNet meshes (subdivision, Blender sculpting, CloudCompare ground truth) to build a scalable benchmark, and train a Point-MAE-style masked reconstruction network with geometry-aware sampling, iterative reconstruction at test time, and combined point-space and feature-space comparison for anomaly scoring.
Key results
- IMRNet reaches 66.1% mean I-AUROC on Anomaly-ShapeNet across 40 categories.
- IMRNet reaches 72.5% mean I-AUROC on Real3D-AD, above RegAD (70.4%).
- Geometry-aware sampling gives I-AUC 0.66 and P-AUC 0.65, outperforming FPS (0.64/0.62), Voxel (0.62/0.55) and random sampling (0.55/0.61).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use IMRNet instead of RegAD | Real3D-AD anomaly detection · Real3D-AD | I-AUROC: 0.704 -> 0.725 | Helped |
| use IMRNet instead of RegAD | Anomaly-ShapeNet anomaly detection · Anomaly-ShapeNet | I-AUROC: 0.572 -> 0.661 | Helped |
| add geometry-aware point-cloud sampling instead of random sampling (ablation) | IMRNet · Anomaly-ShapeNet | I-AUC: 0.55 -> 0.66 | Helped |
| add geometry-aware point-cloud sampling instead of random sampling (ablation) | IMRNet · Anomaly-ShapeNet | P-AUC: 0.61 -> 0.65 | Helped |
| add geometry-aware point-cloud sampling instead of farthest point sampling (ablation) | IMRNet · Anomaly-ShapeNet | I-AUC: 0.64 -> 0.66 | Helped |
| add geometry-aware point-cloud sampling instead of farthest point sampling (ablation) | IMRNet · Anomaly-ShapeNet | P-AUC: 0.62 -> 0.65 | Helped |
| add geometry-aware point-cloud sampling instead of voxel down-sampling (ablation) | IMRNet · Anomaly-ShapeNet | I-AUC: 0.62 -> 0.66 | Helped |
| add geometry-aware point-cloud sampling instead of voxel down-sampling (ablation) | IMRNet · Anomaly-ShapeNet | P-AUC: 0.55 -> 0.65 | Helped |
Surprising findings
- Random sampling gives a weaker I-AUC (0.55) than farthest point sampling, and voxel down-sampling gives the lowest P-AUC (0.55) among the sampling variants.
- Pretrained feature extractors from other datasets show domain bias, which the authors argue motivates self-supervised features.
Limitations
- Anomaly-ShapeNet is synthetic, so realism relative to real industrial defects is not verified against real scans.
- Only four normal training samples per class, a few-shot setting that may limit generalization.
- Iterative test-time reconstruction and feature-template regularization add inference complexity, and the method's gains over baselines on Real3D-AD are modest on some categories.
Experimental setup
Datasets: ShapeNetCoreV2 ShapeNet-55 Real3D-AD
Benchmarks: Anomaly-ShapeNet Real3D-AD
Metrics: I-AUROC P-AUC
Baselines: BTF(Raw) BTF(FPFH) M3DM PatchCore(FPFH) PatchCore(PointMAE) CPMF RegAD
Builds on: Point-MAE PatchCore Real3D-AD ShapeNet
Closest papers by meaning
- Registration is a Powerful Rotation-Invariance Learner for 3D Anomaly Detection (NeurIPS 2025 · similarity 0.90)
- PO3AD: Predicting Point Offsets toward Better 3D Point Cloud Anomaly Detection (CVPR 2025 · similarity 0.90)
- Towards Generalizable 3D Anomaly Detection via Relational Inconsistency Modeling (NeurIPS 2026 · similarity 0.88)
- Looking 3D: Anomaly Detection with 2D-3D Alignment (CVPR 2024 · similarity 0.87)
- PointAD: Comprehending 3D Anomalies from Points and Pixels for Zero-shot 3D Anomaly Detection (NeurIPS 2024 · similarity 0.86)
- RealNet: A Feature Selection Network with Realistic Synthetic Anomaly for Anomaly Detection (CVPR 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.