Source-linked AI summary
Towards Generalizable 3D Anomaly Detection via Relational Inconsistency Modeling
KunHo Heo, SuYeon Kim, Hayoung Lee, Chanse Oh, MyeongAh Cho
TL;DR
Existing 3D anomaly-detection methods learn normality without explicitly characterizing defects, a challenge in unified and cross-domain settings. GRIM learns defect cues from controlled violations of geometric relations, and reports improvements over prior methods across in-domain and cross-domain settings. Its cluster-relative cue may weaken when a large, coherent anomaly forms its own cluster.
Problem
Normality-centered 3D anomaly detection does not explicitly model defects, and diverse normal distributions blur decision boundaries in unified and cross-domain settings.
Source: §1 · p. 2Method
GRIM uses controlled relational pseudo-anomalies, EGR, and CDM to learn a category-agnostic defect criterion based on geometric inconsistencies among neighboring structures.
Source: §1 · pp. 2–3Results
GRIM reports gains of 7.0/3.5 and 7.4/4.7 in object-/point-level AUROC over prior state-of-the-art methods on Real3D-AD and Anomaly-ShapeNet, respectively, with improvements across in- and cross-domain settings.
Source: §1 · p. 3, §5 · p. 10Takeaways & Limitations
The experiments support relational inconsistency modeling as a promising direction for 3D anomaly detection across in- and cross-domain settings.
Source: §5 · p. 10CDM relies on anomalies being sparse relative to normal structures, so its signal may weaken when a large, coherent anomaly forms a separate cluster.
Source: §6.4 · p. 19
Abstract
from arXiv · showhide
3D anomaly detection (3DAD) aims to identify defective regions in point cloud data, serving as a critical component in industrial inspection systems. Existing methods are normality-centered -- learning the distribution of normal samples and treating deviations as anomalies -- without explicitly modeling what constitutes a defect. This leads to ambiguous decision boundaries with increased false positives and negatives, particularly in unified and cross-domain settings where diverse normal distributions further blur the boundaries. We propose a relational inconsistency modeling framework that characterizes defects as violations of geometric consistency among neighboring structures. Our approach learns category-agnostic defect cues through pseudo-anomalies designed as controlled relational violations, instantiated by two key modules: Edge-aware Graph Refinement (EGR) for encoding geometric relationships among local regions, and Cluster-Deviation Modeling (CDM) for identifying regions that are relationally incompatible within their structural peer group. Extensive experiments on Anomaly-ShapeNet and Real3D-AD demonstrate consistent improvements over prior state-of-the-art methods in both in-domain and cross-domain settings, validating the effectiveness of learning an explicit, relation-based defect criterion for 3D anomaly detection. Project page: https://visualsciencelab-khu.github.io/GRIM_project/.
1 Introduction
3D anomaly detection seeks defective regions in point clouds, but normality-centered methods do not explicitly model defects as violations of geometric consistency. GRIM learns category-agnostic defect cues from controlled relational violations and uses EGR and CDM to model structural relationships and peer-group deviations.
1 Introduction
- Normality-centered approaches can mistake rare valid regions for defects and miss true defects, with diverse normal distributions further blurring boundaries in unified and cross-domain settings.3D defects often appear as geometric inconsistencies among neighboring structures rather than isolated local abnormalities.Source: §1 · pp. 1–2
- GRIM uses pseudo-anomalies as controlled relational violations to learn a category-agnostic defect criterion rather than imitate specific defect appearances.The supervision distinguishes structurally coherent regions from relation-violating ones.Source: §1 · pp. 2–3
- EGR encodes geometric relationships among neighboring regions, while CDM scores each region relative to clusters of structurally similar regions.This combines relation-aware message passing with finer-grained detection of deviations within structural peer groups.Source: §1 · p. 3
- Gains of 7.0/3.5 and 7.4/4.7 in object-/point-level AUROC over prior state-of-the-art methods are reported on Real3D-AD and Anomaly-ShapeNet, respectively.The framework also reports strong generalization under cross-domain evaluation.Source: §1 · p. 3
2 Related Work
Prior 3D anomaly-detection methods estimate normality through feature similarity or reconstruction, while GRIM instead defines defects through relational inconsistency. Its pseudo-anomalies supervise controlled violations of geometric relations rather than mimicking specific defect patterns.
2.1 Point Cloud 3D Anomaly Detection
- Feature-similarity and reconstruction methods identify anomalies through deviations from learned normality, but can misclassify normal variations and true defects in unified and cross-domain settings.Feature-similarity methods rely on memory banks that limit unified-setting scalability; reconstruction methods can over-generalize.Source: §2.1 · p. 3
2.2 Learning with Synthetic Anomalies
- Unlike prior synthetic-anomaly methods that proxy real defects, GRIM uses pseudo-anomalies to supervise controlled violations of geometric relations among normal structures.This supports a category-agnostic defect criterion grounded in relational inconsistency.Source: §2.2 · p. 4
3 Methods
The method trains a unified 3D anomaly detector using controlled relational violations, graph-based modeling of neighboring regions, and deviations from structurally similar peers. Its training combines group-level classification with a loss that encourages anomalous nodes to move away from their cluster centers.
3.1 Problem Statement
- A single model is trained across categories using unlabeled normal point clouds, then predicts object-level anomalies and point-level localization scores.The test set contains normal and anomalous samples with object-level and point-level labels.Source: §3.1 · p. 4
3.2 Pseudo-Anomaly Generation
- Pseudo-anomalies are controlled local perturbations that violate geometric relations, generated as bulges, sinks, or holes to supervise category-agnostic defect cues.These perturbations expose local geometric inconsistency rather than approximating the full distribution of real defects or forming a complete defect taxonomy.Source: §3.2 · p. 4
3.3 Edge-aware Graph Refinement
- Edge-aware Graph Refinement represents local groups as graph nodes and uses geometric edge features and relation-aware message passing to encode compatibility among neighboring regions.Group features come from locally normalized neighborhoods encoded by a pretrained local encoder; edge descriptors include relative displacement, point spread, normal inconsistency, and curvature difference.Source: §3.3 · pp. 4–5
3.4 Cluster-Deviation Modeling
- Cluster-Deviation Modeling clusters refined regional features and identifies nodes whose cosine-distance deviations from their assigned cluster centers are large relative to their peers.At inference, clustering is applied to each test sample; deviation embeddings are concatenated with node features for a shared binary classifier.Source: §3.4 · p. 6
3.5 Loss Function
- Training lifts point labels to groups, applies cluster-balanced classification, and penalizes anomalous nodes that remain too close to their assigned cluster centers.The overall objective is a weighted sum of the classification and deviation losses, with a predefined margin for the deviation penalty.Source: §3.5 · pp. 6–7
4 Experiments
Experiments compare GRIM with category-specific and unified baselines on two 3DAD benchmarks, evaluating object- and point-level AUROC in in-domain and cross-domain settings.
4.1 Experiment Settings
- The benchmarks are Anomaly-ShapeNet, with 1,600 samples across 40 categories, and Real3D-AD, with 12 categories; both provide four normal training samples per category.Real3D-AD uses single-view test scans despite 360° training captures; evaluation reports object- and point-level AUROC.Source: §4.1 · pp. 7–8
- The comparison includes category-specific baselines and MC3D-AD [6], a unified model trained jointly across categories.Baseline results come from the respective papers or are reproduced using publicly available code.Source: §4.1 · p. 8
4.2 Quantitative Results
- GRIM gains 7.4 O-AUROC over PASDF [42] and 4.7 P-AUROC over PO3AD [33] on Anomaly-ShapeNet, and 7.0 O-AUROC over PASDF and 3.5 P-AUROC over Reg2Inv [36] on Real3D-AD.It is trained as a single unified model yet surpasses all category-specific baselines.Source: §4.2 · p. 8
- Without fine-tuning, GRIM beats MC3D-AD [6] by 28.2 O-AUROC and 51.3 P-AUROC from Anomaly-ShapeNet to Real3D-AD; its O/P-AUROC drops are 3.6/1.7 versus 22.8/38.5.The reverse transfer direction shows a similar trend.Source: §4.2 · p. 8
4.3 Qualitative Results
- In cross-domain visuals, MC3D-AD [6] misclassifies most points as anomalous, while GRIM localizes anomalies precisely in both settings.Figure 4 shows comparisons across object categories; red regions mark detected anomalies.Source: Figure 4 · p. 8, §4.3 · p. 9
4.4 Discussion
- On Anomaly-ShapeNet, anomalous features average 0.73 distance to cluster centers versus 0.48 for all features, supporting CDM's deviation cue.Figure 5 compares average distances between features and cluster centers.Source: §4.4 · p. 9, Figure 5 · p. 9
- EGR adds 28.3 in-domain P-AUROC; CDM adds 22.1 cross-domain P-AUROC, and the combined components yield gains across all metrics.The ablation study reports O-AUROC and P-AUROC on Anomaly-ShapeNet and Real3D-AD.Source: §4.4 · p. 9, Table 3 · p. 9
5 Conclusion
GRIM models 3D defects as relational inconsistencies, using EGR and CDM with pseudo-anomaly supervision to learn a category-agnostic criterion. Experiments show improvements over prior methods in in-domain and cross-domain settings.
5 Conclusion
- GRIM combines EGR and CDM to learn an explicit, category-agnostic defect criterion through pseudo-anomaly supervision, improving over prior methods across in-domain and cross-domain settings.The framework treats defects as violations of geometric consistency among neighboring structures rather than mere deviations from normality.Source: §5 · p. 10
6 Appendix
The appendix details pseudo-anomaly construction and implementation, then examines generalization, component behavior, efficiency, and limitations. Supplementary tables and figures extend the quantitative and qualitative comparisons.
6.1 Pseudo-Anomaly Generation Details
- Pseudo-anomalies locally perturb a small patch, with bulge, sink, and hole primitives and a point-level mask marking modified regions.The patch is formed from nearest neighbors of a random seed; approximately 1% of object points are modified. Random rotation is applied afterward to encourage pose-agnostic learning.Source: §6.1 · pp. 14–15, Figure 6 · p. 15
6.2 Implementation Details
- Implementation uses a frozen PointMAE feature extractor, an EGR graph over group centroids, and CDM clustering, trained on category-merged data with pseudo-anomalies.At inference, node anomaly scores are propagated to points, smoothed, and max-pooled for the object score; evaluation normalization is applied across samples within each category.Source: §6.2 · pp. 15–16
6.3 Additional Discussion
- CDM performance generally improves with more clusters, while increasing λdev helps in-domain performance but can hurt cross-domain generalization beyond 1.5.The reported clustering trend in cross-domain evaluation holds up to R = 64; larger deviation penalties can tailor features too closely to the training distribution.Source: §6.3 · p. 16, Table 6 · p. 17, Table 5 · p. 17
- Inference takes 0.35s per sample on Anomaly-ShapeNet, more than twice as fast as MC3D-AD, with fewer parameters than that unified baseline.AIT was measured over 1,312 test samples on a single NVIDIA RTX 3090 GPU.Source: §6.3 · p. 17, Table 7 · p. 17
- Unified training reduced PO3AD and PASDF performance; removing either CDM deviation component also reduced results, while the method generalized to unsynthesized defects.The ablation found the largest cross-domain decrease when the deviation embedding was removed and the lowest performance when both deviation components were removed. Broken, bending, and scratch were not explicitly synthesized during training.Source: §6.3 · pp. 17–18, Table 10 · p. 18, Table 11 · p. 19
- Our method outperforms both additional unified cross-domain baselines in object-level detection and point-level localization.The comparison applies models trained on Anomaly-ShapeNet directly to Real3D-AD without fine-tuning or domain-specific adaptation.Source: §6.3 · p. 18, Table 12 · p. 19
6.4 Limitations and Future Works
- The framework uses point clouds alone; CDM’s deviation cue may weaken when a large, coherent anomaly forms its own cluster.The authors identify RGB-D cues such as surface color variation and subtle depth discontinuities as possible complementary inputs, and note that EGR features alone do not establish robustness to large coherent anomalies.Source: §6.4 · p. 19
6.5 Additional Quantitative Results / 6.6 Additional Qualitative Results
- Supplementary tables provide per-category AUROC and AUPR results across in-domain and both cross-domain directions; additional figures show in-domain and cross-domain qualitative comparisons.The cross-domain qualitative comparisons use models trained on Anomaly-ShapeNet and tested on Real3D-AD.Source: §6.5 · p. 19, §6.6 · p. 19