Research map / Label-efficient and robust learning / OOD and anomaly detection

Towards Generalizable 3D Anomaly Detection via Relational Inconsistency Modeling

NeurIPS 2026 · Label-efficient and robust learning › OOD and anomaly detection · approach: Unified multi-class anomaly detection · 0 citations (OpenAlex)

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 GRIM, a unified point-cloud 3D anomaly detection framework that learns category-agnostic defect cues via relational-inconsistency pseudo-anomalies, using Edge-aware Graph Refinement and Cluster-Deviation Modeling.

Problem

Point-cloud 3D anomaly detection is normality-centered, yielding ambiguous decision boundaries and high false positive/negative rates, especially in unified (multi-category) and cross-domain settings.

Approach

Defects are modeled as violations of geometric consistency among neighboring structures: pseudo-anomalies (bulge, sink, hole) supervise a graph over local point groups with edge-aware message passing (EGR), and a k-means cluster-relative cosine deviation feature (CDM) scores regions against their structural peer group.

Key results

• Unified GRIM reaches 97.4 O-AUROC / 94.5 P-AUROC on Anomaly-ShapeNet and 87.2 / 91.3 on Real3D-AD, outperforming prior SOTA (PASDF, PO3AD, Reg2Inv) and unified MC3D-AD.
• Cross-domain Anomaly-ShapeNet to Real3D-AD yields 83.6 O-AUROC / 89.6 P-AUROC, versus MC3D-AD's 55.4 / 38.3.
• Ablation shows EGR and CDM are complementary; removing both drops cross-domain P-AUROC from 89.6 to 42.8 (baseline).

Surprising findings

• Category-specific PO3AD and PASDF degrade substantially when retrained under unified training (e.g., PASDF O-AUROC on Real3D-AD falls from 80.2 to 64.7).
• Hole-only pseudo-anomaly supervision achieves nearly the same Real3D-AD performance as all three primitives (87.1 vs 87.2 O-AUROC).

Limitations

• Relies on anomalies being sparse relative to normal structures; large coherent anomalies may form their own cluster and weaken the CDM signal.
• Operates only on point clouds; no RGB or depth fusion explored.
• Pseudo-anomaly types are not a complete defect taxonomy; increasing anomaly diversity did not monotonically help.

Experimental setup

Datasets: Anomaly-ShapeNet Real3D-AD ModelNet40 (Point-MAE pretraining)

Benchmarks: Anomaly-ShapeNet Real3D-AD

Metrics: O-AUROC P-AUROC

Baselines: BTF M3DM PatchCore CPMF IMRNet Reg3D-AD Group3AD R3D-AD ISMP PO3AD PASDF Reg2Inv

Builds on: Point-MAE PO3AD MC3D-AD FPS k-means++

Closest papers by meaning

• Registration is a Powerful Rotation-Invariance Learner for 3D Anomaly Detection (NeurIPS 2025 · similarity 0.88)
• Towards Scalable 3D Anomaly Detection and Localization: A Benchmark via 3D Anomaly Synthesis and A Self-Supervised Learning Network (CVPR 2024 · similarity 0.88)
• Looking 3D: Anomaly Detection with 2D-3D Alignment (CVPR 2024 · similarity 0.88)
• 3CAD: A Large-Scale Real-World 3C Product Dataset for Unsupervised Anomaly Detection (AAAI 2025 · similarity 0.87)
• PO3AD: Predicting Point Offsets toward Better 3D Point Cloud Anomaly Detection (CVPR 2025 · similarity 0.86)
• Real-IAD D3: A Real-World 2D/Pseudo-3D/3D Dataset for Industrial Anomaly Detection (CVPR 2025 · similarity 0.86)