Towards Generalizable 3D Anomaly Detection via Relational Inconsistency Modeling

KunHo Heo; SuYeon Kim; Hayoung Lee; Chanse Oh; MyeongAh Cho

cs.CVSep 28, 2026

Towards Generalizable 3D Anomaly Detection via Relational Inconsistency Modeling

Authors: KunHo Heo, SuYeon Kim, Hayoung Lee, Chanse Oh, MyeongAh Cho

Organizations: Kyung Hee University

Abstract

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/.

Figures & tables

Figure 1: Comparison of previous normality-centered 3DAD methods and our relation-based defect learning framework. (a) Previous methods (e.g., MC3D-AD) rely on normal feature distributions, often resulting in false positives and false negatives. (b) Our method uses pseudo-anomalies as controlled violations of geometric relations among normal structures. By distinguishing relation-violating pseudo-anomalies from normal regions, the model produces more discriminative point-level anomaly scores, reducing both false positives and false negatives.

Figure 2: Comparison of False Positive Rate (FPR) and False Negative Rate (FNR) between two unified 3DAD models, MC3D-AD and our method.

Figure 3: Overview of the proposed method. Given grouped features from an input sample as nodes, we construct edge features encoding geometric relationships between local groups, and perform Edge-aware Graph Refinement (EGR) to obtain refined node features. The refined features are then passed through Cluster-Deviation Modeling (CDM), which captures relative deviations within clusters, and fed into the classifier together with the features.

Method		Anomaly-ShapeNet		Real3D-AD
Method		O-AUROC( $\uparrow$ )	P-AUROC( $\uparrow$ )	O-AUROC( $\uparrow$ )	P-AUROC( $\uparrow$ )
Category- Specific	BTF(Raw) $({\small{\color[rgb]{0.2344,0.2734,0.4492}\text{CVPR'23}}})$ Horwitz and Hoshen (2023)	49.3	55.0	63.5	57.1
	BTF(FPFH) Horwitz and Hoshen (2023)	52.8	62.8	60.3	73.3
	M3DM $({\small{\color[rgb]{0.2344,0.2734,0.4492}\text{CVPR'23}}})$ Wang et al. (2023)	55.2	61.6	59.4	62.0
	PatchCore(FPFH) $({\small{\color[rgb]{0.2344,0.2734,0.4492}\text{CVPR'22}}})$ Roth et al. (2022)	56.8	58.0	59.3	68.2
	PatchCore(PointMAE) Roth et al. (2022)	56.2	57.7	59.4	62.0

Table 1: Quantitative comparison of mean AUROC (%) at both the object level and point level on Anomaly-ShapeNet and Real3D-AD. The best and second-best results are highlighted in bold and underline , respectively. Per-category results are provided in the supplementary material.

Method	Anomaly-ShapeNet $\rightarrow$ Real3D-AD				Real3D-AD $\rightarrow$ Anomaly-ShapeNet
Method	O-AUROC( $\uparrow$ )	$\Delta$ ( $\downarrow$ )	P-AUROC( $\uparrow$ )	$\Delta$ ( $\downarrow$ )	O-AUROC( $\uparrow$ )	$\Delta$ ( $\downarrow$ )	P-AUROC( $\uparrow$ )	$\Delta$ ( $\downarrow$ )
MC3D-AD	55.4 (78.2)	22.8	38.3 (76.8)	38.5	78.3 (84.2)	5.9	48.8 (75.9)	27.1
GRIM (Ours)	83.6 (87.2)	3.6	89.6 (91.3)	1.7	91.7 (97.4)	5.7	78.9 (94.5)	15.6

Table 2: Quantitative comparison of mean AUROC (%) under the cross-domain setting. In each

\rightarrow

pair, the left denotes the training set and the right denotes the test set. Values in ( ) indicate in-domain performance, and

\Delta

denotes the performance gap between in-domain and cross-domain results.

Figure 4: Qualitative comparison across various object categories, showing in-domain results (trained and tested on Anomaly-ShapeNet) and cross-domain results (trained on Anomaly-ShapeNet and tested on Real3D-AD). Red colored regions indicate detected anomalies.

Figure 5: Average distances between features and cluster centers.

EGR	CDM	In-Domain (A $\rightarrow$ A)		Cross-Domain (A $\rightarrow$ R)
EGR	CDM	O-R( $\uparrow$ )	P-R( $\uparrow$ )	O-R( $\uparrow$ )	P-R( $\uparrow$ )
		89.5	58.1	56.9	42.8
✓		94.3	86.4	59.4	53.0
	✓	96.6	79.8	62.3	64.9
✓	✓	97.4	94.5	83.6	89.6

Table 3: Ablation studies on the proposed components. O-AUROC (O-R) and P-AUROC (P-R) are reported on Anomaly-ShapeNet (A) and Real3D-AD (R).

Bulge	Sink	Hole	In-Domain (A $\rightarrow$ A)		Cross-Domain (A $\rightarrow$ R)
Bulge	Sink	Hole	O-R( $\uparrow$ )	P-R( $\uparrow$ )	O-R( $\uparrow$ )	P-R( $\uparrow$ )
✓			96.9	94.1	77.8	88.0
	✓		98.5	95.9	81.2	81.8
		✓	95.9	93.9	78.4	86.8
✓	✓		96.9	94.1	81.8	89.0
✓		✓	98.6	96.0	80.3	87.9

Table 4: Ablation studies on anomaly types.

Figure 6: Visualization of pseudo-anomalies generated on Anomaly-ShapeNet and Real3D-AD. Gray points represent object point clouds, and red points denote the perturbed regions used as pseudo-anomaly masks.

Figure 7: t-SNE visualization of clustering results in CDM. Nodes of the same color belong to the same cluster.

$R$	In-Domain (A $\rightarrow$ A)		Cross-Domain (A $\rightarrow$ R)
$R$	O-AUROC( $\uparrow$ )	P-AUROC( $\uparrow$ )	O-AUROC( $\uparrow$ )	P-AUROC( $\uparrow$ )
16	96.1	89.9	63.8	73.3
32	96.8	92.5	68.4	79.0
64	97.4	94.5	83.6	89.6

Table 5: Ablation studies on the number of clusters

R

on Anomaly-ShapeNet (A) and Real3D-AD (R).

$\lambda_{\mathrm{dev}}$	In-Domain (A $\rightarrow$ A)		Cross-Domain (A $\rightarrow$ R)
$\lambda_{\mathrm{dev}}$	O-AUROC( $\uparrow$ )	P-AUROC( $\uparrow$ )	O-AUROC( $\uparrow$ )	P-AUROC( $\uparrow$ )
0.5	96.5	93.1	81.6	88.4
1.0	97.3	94.5	81.8	86.7
1.5	97.4	94.5	83.6	89.6

Table 6: Ablation studies on the weighting coefficient

\lambda_{\mathrm{dev}}

Method		AIT $\downarrow$	FPS $\uparrow$	#Params $\downarrow$	FLOPs $\downarrow$
Category- Specific	PO3AD	0.31s	3.23	433.6M	399
	PASDF	0.92s	1.09	15.9M	624
	Reg2Inv	0.78s	1.28	85.3M	815
Unified	MC3D-AD	0.84s	1.19	34.2M	417
Unified	GRIM (Ours)	0.35s	2.86	34.0M	371

Table 7: Comparison of computational complexity.

Metric	Bulge	Concavity	Hole	Broken	Bending	Scratch
O-AUROC( $\uparrow$ )	96.1	95.9	97.0	96.9	96.6	97.7
P-AUROC( $\uparrow$ )	94.2	94.6	93.3	93.8	94.8	95.3

Table 8: Defect-type breakdown on Anomaly-ShapeNet.

Training pseudo-anomalies	O-AUROC $\uparrow$	P-AUROC $\uparrow$
Hole	87.1	89.2
Bulge, Sink, Hole	87.2	91.3

Table 9: Cross-type supervision results on Real3D-AD.

Configuration	In-Domain ( $A\rightarrow A$ )		Cross-Domain ( $A\rightarrow R$ )
Configuration	O-AUROC $\uparrow$	P-AUROC $\uparrow$	O-AUROC $\uparrow$	P-AUROC $\uparrow$
Full CDM	97.4	94.5	83.6	89.6
w/o $\mathcal{L}_{\mathrm{dev}}$	96.6	92.1	79.3	84.1
w/o deviation emb.	94.9	88.7	68.5	66.8
w/o both	94.3	86.4	59.4	53.0

Table 10: Ablation of the deviation loss and deviation embedding in CDM.

	Real3D-AD		Anomaly-ShapeNet
Method	O-AUROC $\uparrow$	P-AUROC $\uparrow$	O-AUROC $\uparrow$	P-AUROC $\uparrow$
PO3AD (C)	76.5	65.0	83.9	89.8
PO3AD (U)	54.7	62.2	72.3	75.1
PASDF (C)	80.2	74.5	90.0	89.7
PASDF (U)	64.7	58.4	77.5	73.6

Table 11: Comparison between category-specific and unified training on Real3D-AD and Anomaly-ShapeNet.(C) denotes the category-specific setting and (U) denotes the unified setting.

Method	O-AUROC $\uparrow$	P-AUROC $\uparrow$
PO3AD (U)	52.2	48.7
PASDF (U)	50.8	43.4
Ours	83.6	89.6

Table 12: Additional unified cross-domain comparison. (U) denotes the unified setting.

O-AUROC( $\uparrow$ )
Method		ashtray0	bag0	bottle0	bottle1	bottle3	bowl0	bowl1	bowl2	bowl3	bowl4	bowl5	bucket0	bucket1	cap0
Category- Specific	BTF(Raw)	57.8	41.0	59.7	51.0	56.8	56.4	26.4	52.5	38.5	66.4	41.7	61.7	32.1	66.8
	BTF(FPFH)	42.0	54.6	34.4	54.6	32.2	50.9	66.8	51.0	49.0	60.9	69.9	40.1	63.3	61.8
	M3DM	57.7	53.7	57.4	63.7	54.1	63.4	66.3	68.4	61.7	46.4	40.9	30.9	50.1	55.7
	PatchCore(FPFH)	58.7	57.1	60.4	66.7	57.2	50.4	63.9	61.5	53.7	49.4	55.8	46.9	55.1	58.0
	PatchCore(PointMAE)	59.1	60.1	51.3	60.1	65.0	52.3	62.9	45.8	57.9	50.1	59.3	59.3	56.1	58.9

Table 13: Comparison of object-level AUROC (%) across all categories on Anomaly-ShapeNet.

P-AUROC( $\uparrow$ )
Method		ashtray0	bag0	bottle0	bottle1	bottle3	bowl0	bowl1	bowl2	bowl3	bowl4	bowl5	bucket0	bucket1
Category- Specific	BTF(Raw)	51.2	43.0	55.1	49.1	72.0	52.4	46.4	42.6	68.5	56.3	51.7	61.7	68.6
	BTF(FPFH)	62.4	74.6	64.1	54.9	62.2	71.0	76.8	51.8	59.0	67.9	69.9	40.1	63.3
	M3DM	57.7	63.7	66.3	63.7	53.2	65.8	66.3	69.4	65.7	62.4	48.9	69.8	69.9
	PatchCore(FPFH)	59.7	57.4	65.4	68.7	51.2	52.4	53.1	62.5	32.7	72.0	35.8	45.9	57.1
	PatchCore(PointMAE)	49.5	67.4	55.3	60.6	65.3	52.7	52.4	51.5	58.1	50.1	56.2	58.6	57.4

Table 14: Comparison of point-level AUROC (%) across all categories on Anomaly-ShapeNet.

O-AUROC( $\uparrow$ )
Method		Airplane	Car	Candybar	Chicken	Diamond	Duck	Fish	Gemstone	Seahorse	Shell	Starfish	Toffees	Mean
Category- Specific	BTF(Raw)	73.0	64.7	53.9	78.9	70.7	69.1	60.2	68.6	59.6	39.6	53.0	70.3	63.5
	BTF(FPFH)	52.0	56.0	63.0	43.2	54.5	78.4	54.9	64.8	77.9	75.4	57.5	46.2	60.3
	M3DM	43.4	54.1	55.2	68.3	60.2	43.3	54.0	64.4	49.5	69.4	55.1	45.0	59.4
	PatchCore(FPFH)	88.2	59.0	54.1	83.7	57.4	54.6	67.5	37.0	50.5	58.9	44.1	56.5	59.3
	PatchCore(PointMAE)	72.6	49.8	66.3	82.7	78.3	48.9	63.0	37.4	53.9	50.1	51.9	58.5	59.4

Table 15: Comparison of object-level AUROC (%) across all categories on Real3D-AD.

P-AUROC( $\uparrow$ )
Method		Airplane	Car	Candybar	Chicken	Diamond	Duck	Fish	Gemstone	Seahorse	Shell	Starfish	Toffees	Mean
Category- Specific	BTF(Raw)	56.4	64.7	73.5	60.9	56.3	60.1	51.4	59.7	52.0	48.9	39.2	62.3	57.1
	BTF(FPFH)	73.8	70.8	86.4	73.5	88.2	87.5	70.9	89.1	51.2	57.1	50.1	81.5	73.3
	M3DM	54.7	60.2	67.9	67.8	60.8	66.7	60.6	67.4	56.0	73.8	53.2	68.2	62.0
	PatchCore(FPFH)	56.2	75.4	78.0	42.9	82.8	26.4	82.9	91.0	73.9	73.9	60.6	74.7	68.2
	PatchCore(PointMAE)	56.9	60.9	62.7	72.9	71.8	52.8	71.7	44.4	63.3	70.9	58.0	58.0	62.0

Table 16: Comparison of point-level AUROC (%) across all categories on Real3D-AD.

O-AUROC( $\uparrow$ )
Method	Airplane	Car	Candybar	Chicken	Diamond	Duck	Fish	Gemstone	Seahorse	Shell	Starfish	Toffees	Mean
MC3D-AD	42.8	48.2	58.8	50.5	67.0	63.3	49.5	57.1	57.8	59.4	65.6	45.4	55.4
GRIM (Ours)	78.0	70.2	76.4	81.2	96.2	85.9	81.6	81.0	92.8	97.1	88.0	74.6	83.6

Table 17: Comparison of object-level AUROC (%) across all categories on Real3D-AD under the cross-domain setting. Models are trained on Anomaly-ShapeNet and tested on Real3D-AD.

P-AUROC( $\uparrow$ )
Method	Airplane	Car	Candybar	Chicken	Diamond	Duck	Fish	Gemstone	Seahorse	Shell	Starfish	Toffees	Mean
MC3D-AD	38.5	25.0	32.1	44.1	37.4	36.5	35.9	41.2	36.3	53.8	50.1	29.1	38.3
GRIM (Ours)	76.3	87.0	92.0	84.6	93.9	90.1	95.2	92.2	86.1	97.3	89.6	91.3	89.6

Table 18: Comparison of point-level AUROC (%) across all categories on Real3D-AD under the cross-domain setting. Models are trained on Anomaly-ShapeNet and tested on Real3D-AD.

O-AUROC( $\uparrow$ )
Method	ashtray0	bag0	bottle0	bottle1	bottle3	bowl0	bowl1	bowl2	bowl3	bowl4	bowl5	bucket0	bucket1	cap0	cap3	cap4	cap5	cup0	cup1	eraser0
MC3D-AD	62.9	68.1	77.6	80.0	78.1	97.8	73.7	75.2	74.1	78.5	76.1	74.6	75.2	61.9	89.8	87.0	80.4	74.8	86.2	80.0
GRIM (Ours)	96.2	79.0	85.7	82.8	99.7	97.0	91.1	100.0	93.7	91.5	95.4	94.0	85.1	87.0	100.0	100.0	95.4	91.0	91.4	80.0

Table 19: Comparison of object-level AUROC (%) across all categories on Anomaly-ShapeNet under the cross-domain setting. Models are trained on Real3D-AD and tested on Anomaly-ShapeNet.

P-AUROC( $\uparrow$ )
Method	ashtray0	bag0	bottle0	bottle1	bottle3	bowl0	bowl1	bowl2	bowl3	bowl4	bowl5	bucket0	bucket1
MC3D-AD	48.6	51.3	56.2	62.7	48.5	49.5	43.4	48.6	43.1	50.3	43.6	55.3	47.3
GRIM (Ours)	82.1	85.6	86.5	78.6	84.9	75.6	61.7	84.6	77.0	79.3	81.6	77.1	80.8

Table 20: Comparison of point-level AUROC (%) across all categories on Anomaly-ShapeNet under the cross-domain setting. Models are trained on Real3D-AD and tested on Anomaly-ShapeNet.

O-AUPR( $\uparrow$ )
Method		ashtray0	bag0	bottle0	bottle1	bottle3	bowl0	bowl1	bowl2	bowl3	bowl4	bowl5	bucket0	bucket1
Category- Specific	BTF(Raw)	57.8	45.8	46.6	57.3	54.3	58.8	46.4	57.6	65.4	60.1	61.5	65.2	62.0
	BTF(FPFH)	65.1	55.1	64.4	62.5	60.2	57.6	64.8	51.5	49.9	63.2	69.9	48.3	64.8
	M3DM	63.2	64.2	76.3	67.4	45.1	52.5	51.5	63.0	63.5	57.1	60.1	60.9	50.7
	PatchCore(FPFH)	44.5	60.8	61.5	67.7	57.9	54.8	54.5	61.1	62.0	57.5	54.1	60.4	56.5
	PatchCore(PointMAE)	67.9	60.1	54.5	64.5	65.1	56.2	61.1	45.6	55.6	60.1	58.5	54.1	64.2

Table 21: Comparison of object-level AUPR (%) across all categories on Anomaly-ShapeNet.

Figure 8: Additional in-domain qualitative comparison with MC3D-AD on Anomaly-ShapeNet.

Figure 9: Additional in-domain qualitative comparison with MC3D-AD on Anomaly-ShapeNet.

Figure 10: Additional in-domain qualitative comparison with MC3D-AD on Anomaly-ShapeNet.

Figure 11: Additional in-domain qualitative comparison with MC3D-AD on Anomaly-ShapeNet.

Figure 12: Additional cross-domain qualitative comparison with MC3D-AD. Models are trained on Anomaly-ShapeNet and tested on Real3D-AD.

Figure 13: Additional cross-domain qualitative comparison with MC3D-AD. Models are trained on Anomaly-ShapeNet and tested on Real3D-AD.

Explore similar work

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
Physics-inspired Pseudo Anomaly Generation and Prototype Feature Guidance for 3D Anomaly Detection
Jul 12, 2026Jian Ning, Qin Zou, Linchun Wu +4Point Clouds 3D Object Detection
School of Computer Science, Wuhan University, Wuhan 430072, Hubei, China · Guangdong Laboratory of Artificial Intelligence and Digital Economy (SZ), Shenzhen 518060, Guangdong, China · School of Information and Software Engineering, East China Jiaotong University, Nanchang 330013, Jiangxi, China +1
M2P-AD: Memory-to-Prototype Learning with Boundary-aware Score Refinement for 3D Anomaly Detection
Jul 15, 2026Seyoung Jeong, Jong Pil Yun, Sang Jun Lee3D Object Detection Boundaries
Jeonbuk National University, Jeonju, Republic of Korea · Manufacturing AI Research Center, Korea Institute of Industrial Technology (KITECH), Incheon, Republic of Korea

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

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

Ali Balapour, Faraz Hach

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

3D Object Detection Point Clouds Unsupervised Modular Framework

Open paper AskPDF arXiv

Detecting and localizing defects in 3D point clouds is challenging because abnormal samples are scarce and diverse, while training is often limited to normal data. We propose Anomaly Factory 3D (AF3AD), a modular framework that synthesizes diverse pseudo-anomalies from normal point clouds to expand the training data for unsupervised 3D anomaly detection methods that rely on pseudo-anomalies. AF3AD uses a center-conditioned parametric deformation model defined in local PCA frames, with kernel-controlled spatial falloff, anisotropy, directional gating, and normal/tangential displacement fields, enabling a broad set of geometric defect presets. We demonstrate its ease-of-use and effectiveness by integrating AF3AD with an offset-prediction detector and a reconstruction-based anomaly detection method, showing that AF3AD transfers across detection paradigms. Experiments on AnomalyShapeNet and Real3D-AD show consistent improvements in object- and point-level detection and localization, supported by ablations on preset groups and robustness under noise. AF3AD is designed as a standalone synthesis tool to facilitate adoption across different 3D anomaly detection paradigms. Code is available at github.com/vpc-ccg/AF3AD.

Abstract

Explore similar work

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

Physics-inspired Pseudo Anomaly Generation and Prototype Feature Guidance for 3D Anomaly Detection

M2P-AD: Memory-to-Prototype Learning with Boundary-aware Score Refinement for 3D Anomaly Detection