Beyond Geometry: Benchmarking and Consistency Reasoning for 3D Logical Anomaly Detection

Zhiqiang Qin; He Xie; Junfei Yi; Yang Yang; Hao Wang; Yunkang Cao; Hui Zhang; Yaonan Wang

cs.CVSep 28, 2026

Beyond Geometry: Benchmarking and Consistency Reasoning for 3D Logical Anomaly Detection

Authors: Zhiqiang Qin, He Xie, Junfei Yi, Yang Yang, Hao Wang, Yunkang Cao, Hui Zhang, Yaonan Wang

Organizations: School of Artificial Intelligence and Robotics, Hunan University · National Engineering Research Center of Robot Visual Perception and Control Technology, Hunan University

Abstract

Existing 3D industrial anomaly detection mainly targets local geometric deviations. In contrast, many industrial anomalies violate object-level design or assembly rules, which we define as 3D logical anomalies. To address these challenges, we introduce the Industrial Logical Anomaly Detection Dataset (ILGAD), the first scalable benchmark dedicated to logical anomalies in industrial point clouds. ILGAD contains 2,774 samples from 15 categories with point-level annotations and covers existence, specification, pose, and assembly-state errors. To detect such 3D logical anomalies, we propose a consistency reasoning framework that assesses whether local geometry, structure coverage, and spatial relations conform to the normal design. The framework detects geometric changes, unsupported expected structures, and abnormal local arrangements. Experiments on ILGAD, Anomaly-ShapeNet, and IEC3D demonstrate superior object-level detection and point-level localization, showing that the framework effectively detects logical anomalies and generalizes to conventional geometric defects.

Figures & tables

Figure 1: Comparison between representative existing 3D anomaly detection datasets and ILGAD. Existing benchmarks mainly focus on geometric anomalies, whereas ILGAD covers existence, specification, pose, and assembly-state errors. Red markings indicate anomalies, and the insets show the corresponding normal structures.

Dataset	Logical	Rule-guided	GT	Asm. Obj.
MVTec 3D-AD	$\times$	$\times$	✓	$\times$
Real3D-AD	$\times$	$\times$	✓	$\times$
Anomaly-ShapeNet	$\times$	$\times$	✓	$\times$
IEC3D	$\times$	$\times$	✓	$\times$
ILGAD	✓	✓	✓	✓

Table 1: Comparison of representative 3D anomaly detection datasets. Rule-guided indicates that anomalous CAD models are manually constructed by violating predefined structural or assembly requirements. Asm. Obj. indicates whether the dataset contains multi-component assembled objects.

Figure 2: Overview of the Consistency Reasoning Framework. After canonical reference alignment, three parallel components evaluate local geometric consistency, structure coverage consistency, and spatial relation consistency. The resulting scores are propagated to visible test points and fused to produce a point-level anomaly map and an object-level anomaly score.

Category	PC-FPFH	PC-MAE	BTF-Raw	BTF-FPFH	Reg3D-AD	PO3AD	Template3D	Simple3D	Ours
	CVPR’22	CVPR’22	CVPRW’23	CVPRW’23	NeurIPS’23	CVPR’25	IJCAI’25	AAAI’26
CLB	90.61 /72.49	42.07/55.69	47.20/47.30	58.00/70.20	83.85/64.81	57.34/65.48	72.83/ 95.75	65.00/68.90	100.00 / 99.25
BC	82.05 /78.26	61.05/60.03	54.00/56.00	52.50/58.00	78.12 /81.61	47.45/65.89	52.30/ 90.73	62.50/63.90	77.95/ 94.99
WH	65.43/77.15	47.97/64.16	55.60/58.00	46.20/49.30	65.85/83.41	67.95 /79.61	68.05 / 84.82	58.00/54.30	61.52/ 86.68
BCA	81.55 /82.47	50.53/71.15	46.90/61.10	47.50/60.70	67.53/73.44	60.68/74.84	53.25/ 95.02	78.00/64.60	99.72 / 98.95
CDB	88.49/79.14	72.75/76.05	56.30/45.60	53.60/63.80	81.81/84.12	64.31/67.67	91.11 / 93.18	73.30/55.60	99.78 / 94.98

Table 2: Quantitative results on ILGAD. The results are reported as O-ROC%/P-ROC%. The best performance is in bold , and the second best is underlined .

Figure 3: Qualitative anomaly localization results on ILGAD, Anomaly-ShapeNet, and IEC3D. For each dataset, the top row shows the input point clouds, the middle row shows the point-level ground truth, and the bottom row shows the predicted anomaly maps. Red regions indicate anomalous areas or high anomaly responses.

Method	Pub./Year	O-ROC	P-ROC
PC-FPFH	CVPR’22	56.80	58.00
PC-MAE	CVPR’22	56.20	57.70
BTF-Raw	CVPRW’23	49.30	55.00
BTF-FPFH	CVPRW’23	52.80	62.80
M3DM	CVPR’23	55.20	61.60
Reg3D-AD	NeurIPS’23	57.20	66.80

Table 3: Quantitative results on Anomaly-ShapeNet. The results are reported as O-ROC% and P-ROC%. The best performance is in bold , and the second best is underlined .

Method	Pub./Year	O-ROC	P-ROC
BTF-Raw	CVPRW’23	68.93	70.17
BTF-FPFH	CVPRW’23	48.05	53.40
M3DM-PointMAE	CVPR’23	59.65	55.65
M3DM-PointBERT	CVPR’23	58.38	54.17
PC-FPFH	CVPR’22	86.99	64.26
PC-FPFH+Raw	CVPR’22	87.05	66.86

Table 4: Quantitative results on IEC3D. The results are reported as O-ROC% and P-ROC%. The best performance is in bold , and the second best is underlined .

Type	Geo.	Cov.	Rel.	O-ROC	P-ROC
Overall	✓			89.34	89.10
	✓	✓		90.79	96.16
	✓		✓	88.33	89.60
	✓	✓	✓	91.90	96.28
Logical	✓			90.47	90.28
	✓	✓		91.18	96.51

Table 5: Ablation study of the proposed consistency branches on ILGAD.

Figure 4: Visualization of the three consistency responses on representative logical anomalies.

Explore similar work

Towards Generalizable 3D Anomaly Detection via Relational Inconsistency Modeling
Sep 28, 2026KunHo Heo, SuYeon Kim, Hayoung Lee +23D Object Detection
Kyung Hee University
Breaking the Rigid Prior: Towards Articulated 3D Anomaly Detection
Apr 29, 2026Jinye Gan, Bozhong Zheng, Xiaohao Xu +43D Object Detection 3D Models
ShanghaiTech University · China · University of Michigan, Ann Arbor +1
Anomaly Factory 3D: A Modular Framework for Diverse Pseudo-Anomaly Synthesis in Unsupervised 3D Anomaly Detection
Jun 28, 2026Ali Balapour, Faraz Hach3D Object Detection Point Clouds
University of British Columbia, Vancouver, V6T 1Z4 BC, Canada · Vancouver Prostate Centre, M.H. Mohseni Institute of Urologic Sciences, Vancouver, V6H 3ZB BC, Canada · Department of Urologic Sciences, The University of British Columbia, Vancouver, V5Z 1M9 BC, Canada

Sep 28, 2026, cs.CVJ/K move · Enter open · S save

Towards Generalizable 3D Anomaly Detection via Relational Inconsistency Modeling

KunHo Heo, SuYeon Kim, Hayoung Lee, Chanse Oh+1

Kyung Hee University

3D Object Detection

Open paper AskPDF arXiv

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/.

Abstract

Explore similar work

Towards Generalizable 3D Anomaly Detection via Relational Inconsistency Modeling

Breaking the Rigid Prior: Towards Articulated 3D Anomaly Detection

Anomaly Factory 3D: A Modular Framework for Diverse Pseudo-Anomaly Synthesis in Unsupervised 3D Anomaly Detection