A Semantically Disentangled Unified Model for Multi-category 3D Anomaly Detection

SuYeon Kim; Wongyu Lee; MyeongAh Cho

cs.CVMar 26, 2026

A Semantically Disentangled Unified Model for Multi-category 3D Anomaly Detection

Authors: SuYeon Kim, Wongyu Lee, MyeongAh Cho

Organizations: Kyung Hee University, Republic of Korea

Abstract

3D anomaly detection targets the detection and localization of defects in 3D point clouds trained solely on normal data. While a unified model improves scalability by learning across multiple categories, it often suffers from Inter-Category Entanglement (ICE)-where latent features from different categories overlap, causing the model to adopt incorrect semantic priors during reconstruction and ultimately yielding unreliable anomaly scores. To address this issue, we propose the Semantically Disentangled Unified Model for 3D Anomaly Detection, which reconstructs features conditioned on disentangled semantic representations. Our framework consists of three key components: (i) Coarse-to-Fine Global Tokenization for forming instance-level semantic identity, (ii) Category-Conditioned Contrastive Learning for disentangling category semantics, and (iii) a Geometry-Guided Decoder for semantically consistent reconstruction. Extensive experiments on Real3D-AD and Anomaly-ShapeNet demonstrate that our method achieves state-of-the-art for both unified and category-specific models, improving object-level AUROC by 2.8% and 9.1%, respectively, while enhancing the reliability of unified 3D anomaly detection.

Figures & tables

Figure 2: T-SNE visualization and quantitative analysis of semantic disentanglement. MC3D-AD exhibits entangled feature clusters across similar categories ( e.g . , chicken, duck, gemstone ), where samples with low category classification scores are reconstructed under uncertain semantic priors, yielding high reconstruction errors on normal data and increased false positives. Our model instead forms well-separated, semantically aligned manifolds with higher classification scores and lower reconstruction errors, enabling reliable category-conditioned reconstruction.

Figure 3: The overview of the proposed method. Our method consists of two main stages: Semantically Disentangled Representation Learning and Semantically Disentangled Reconstruction. Given an input point cloud, CFGT encodes multi-resolution geometric features into a category-aware global token, C3L disentangles the latent semantics, and GGD reconstructs the object conditioned on these disentangled semantics and geometric priors.

Figure 4: Overview of the Geometry-Guided Decoder (GGD). Geometric priors $\mathbf{B}_{\mathrm{geo}}$ guide the attention mechanism to refine semantic features, and the results are aggregated through a feed-forward network.

O-AUROC( $\uparrow$ )
Method	Airplane	Car	Candybar	Chicken	Diamond	Duck	Fish	Gemstone	Seahorse	Shell	Starfish	Toffees	Mean
Category-Specific Method
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

Table 1: Quantitative comparison of AUROC (%) at the object levels on Real3D-AD across 12 categories. The best and second-best results are highlighted in bold and underline , respectively. Methods in the upper block are trained and evaluated in the single-category setting, while methods in the lower block use a unified model for the multi-category setting across all 12 categories.

O-AUROC( $\uparrow$ )
Method	ashtray0	bag0	bottle0	bottle1	bottle3	bowl0	bowl1	bowl2	bowl3	bowl4	bowl5	bucket0	bucket1	cap0	cap3	cap4	cap5	cup0	cup1	eraser0
Category-Specific Method
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	52.7	46.8	37.3	40.3	52.1	52.5
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	52.2	52.0	58.6	58.6	61.0	71.9
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	42.3	77.7	63.9	53.9	55.6	62.7
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	45.3	75.7	79.0	60.0	58.6	65.7

Table 2: Comparison of object-level AUROC (%) of various methods on the Anomaly-ShapeNet.

Method	Real3D-AD	Anomaly-ShapeNet
Category-Specific Method
BTF(Raw)	57.1	55.0
BTF(FPFH)	73.3	62.8
M3DM	62.0	61.6
PatchCore(FPFH)	68.2	58.0
PatchCore(PointMAE)	62.0	57.7

Table 3: Comparison of mean point-level AUROC (%) on Real3D-AD and Anomaly-ShapeNet datasets. The detailed per-category results are provided in the supplementary.

CFGT	C3L	GGD	O-AUROC ( $\uparrow$ )	P-AUROC ( $\uparrow$ )
			75.1	78.1
✓			76.4	78.4
✓	✓		79.2	79.0
✓	✓	✓	81.0	80.6

Table 4: Ablation studies on the effectiveness of core components.

Figure 5: Qualitative comparison with MC3D-AD across various object categories in Real3D-AD dataset. Red colored regions indicate abnormalities ( e.g . , bulge, sink). MC3D-AD often misses true anomalies or yields false positives, whereas our method delivers more accurate and complete localization aligned with the ground truth.

$\mathcal{L}_{\text{scl}}$	$\mathcal{L}_{\text{cls}}$	$\mathcal{L}_{\text{cos}}$	O-AUROC ( $\uparrow$ )	P-AUROC ( $\uparrow$ )
✓			79.1	78.8
✓	✓		80.0	79.3
✓	✓	✓	81.0	80.6

Table 5: Ablation studies on the contribution of each loss in C3L.

Method	O-AUROC ( $\uparrow$ )	P-AUROC ( $\uparrow$ )
Masking	79.8	79.6
Gating	79.7	78.9
Bias (Ours)	81.0	80.6

Table 6: Comparison of different guidance strategies in GGD.

Method	O-AUROC ( $\uparrow$ )	P-AUROC ( $\uparrow$ )
ACT	78.5	78.5
Max	76.4	75.5
Mean	78.2	78.7
CFGT (Ours)	81.0	80.6

Table 7: Comparison of different global token strategies.

Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

Method	Parameters (M)	FLOPs (G)
MC3D-AD	34.16	439.21
Ours	33.25	508.25

Appendix

Table S1 : Comparison of computational cost.

CFGT	$\mathcal{L}_{\text{scl}}$	$\mathcal{L}_{\text{cos}}$	$\mathcal{L}_{\text{cls}}$	O-AUROC ( $\uparrow$ )	P-AUROC ( $\uparrow$ )
	✓			78.8	79.0
		✓		77.9	78.7
	✓	✓		80.5	78.6
✓	✓			80.0	79.1
✓		✓		79.1	78.8
✓			✓	79.3	79.6

Appendix

Table S2 : Ablation studies on the CFGT and C3L losses.

Ratios	O-AUROC ( $\uparrow$ )	P-AUROC ( $\uparrow$ )
{k, 4k, k/4}	80.3	79.4
{k, 3k, k/3}	80.9	79.8
{k, 1.5k, 2k/3}	79.9	79.8
{k, 2k, k/2}	81.0	80.6

Appendix

Table S3 : Comparison of different coarse-to-fine neighborhood ratios.

Size	O-AUROC ( $\uparrow$ )	P-AUROC ( $\uparrow$ )
{256, 512 , 1024}	77.5	78.5
{64, 128 , 256}	79.9	79.6
{128, 256 , 512}	81.0	80.6

Appendix

Table S4 : Comparison of different base neighborhood sizes

k

N	O-AUROC ( $\uparrow$ )	P-AUROC ( $\uparrow$ )
512	75.7	77.4
2048	77.7	78.4
1024	81.0	80.6

Appendix

Table S5 : Comparison of different numbers of groups.

L	O-AUROC ( $\uparrow$ )	P-AUROC ( $\uparrow$ )
32	79.6	79.8
128	80.3	80.3
10,000	80.5	79.2
64	81.0	80.6

Appendix

Table S6 : Comparison of different buffer sizes in C3L.

$\lambda_{\mathrm{scl}}$	$\lambda_{\mathrm{cos}}$	$\lambda_{\mathrm{cls}}$	O-AUROC ( $\uparrow$ )	P-AUROC ( $\uparrow$ )
0.01	0.001	0.01	79.3	79.6
0.001	0.01	0.01	78.6	78.3
0.001	0.001	0.1	77.0	78.9
0.001	0.001	0.01	81.0	80.6

Appendix

Table S7 : Comparison of different loss weights in C3L.

Method	Accuracy (%)
Mean	78.1
Max	77.3
ACT	97.3

Appendix

Table S8 : Category prediction accuracy with different global representations.

Figure S1 : t-SNE visualization of the learned global tokens on (a) Real3D-AD (12 categories) and (b) Anomaly-ShapeNet (40 categories). The embeddings form well-structured clusters, indicating that the proposed method learns coherent and category-aware global representations across both datasets.

P-AUROC( $\uparrow$ )
Method	Airplane	Car	Candybar	Chicken	Diamond	Duck	Fish	Gemstone	Seahorse	Shell	Starfish	Toffees	Mean
Category-Specific Method
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

Appendix

Table S9 : Comparison of point-level AUROC (%) of various methods on the Real3D-AD dataset.

P-AUROC( $\uparrow$ )
Method	ashtray0	bag0	bottle0	bottle1	bottle3	bowl0	bowl1	bowl2	bowl3	bowl4	bowl5	bucket0	bucket1
Category-Specific Method
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

Appendix

Table S10 : Comparison of point-level AUROC (%) of various methods on the Anomaly-ShapeNet dataset.

O-AUPR( $\uparrow$ )
Method	ashtray0	bag0	bottle0	bottle1	bottle3	bowl0	bowl1	bowl2	bowl3	bowl4	bowl5	bucket0	bucket1
Category-Specific Method
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

Appendix

Table S11 : Comparison of object-level AUPR (%) of various methods on the Anomaly-ShapeNet dataset.

Figure S2 : Additional qualitative comparison with MC3D-AD on the Anomaly-ShapeNet dataset.

Explore similar work

Towards Generalizable 3D Anomaly Detection via Relational Inconsistency Modeling
Sep 28, 2026KunHo Heo, SuYeon Kim, Hayoung Lee +2Point Cloud Anomaly Detection Industrial Anomaly Detection
Kyung Hee University
AT3D-AD: Anomaly Type-Aware 3D Anomaly Detection via Hierarchical Point-Language Alignment
Sep 22, 2026Jingyu Zeng, Haoquan Lu, Can GaoPoint Cloud Anomaly Detection Industrial Anomaly Detection
College of Computer Science and Software Engineering and the Guangdong Key Laboratory of Intelligent Information Processing, Shenzhen University, Shenzhen 518060, China
Align3D-AD: Cross-Modal Feature Alignment and Dual-Prompt Learning for Zero-shot 3D Anomaly Detection
May 7, 2026Letian Bai, Xuanming Cao, Juan Du +1Prompt Learning Cross-Modal Alignment
Smart Manufacturing Thrust, The Hong Kong University of Science and Technology (Guangzhou) · The Hong Kong University of Science and Technology · College of Mechanical and Vehicle Engineering, Hunan University

Sep 28, 2026J/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

Point Cloud Anomaly Detection Industrial Anomaly Detection Anomaly 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

AT3D-AD: Anomaly Type-Aware 3D Anomaly Detection via Hierarchical Point-Language Alignment

Align3D-AD: Cross-Modal Feature Alignment and Dual-Prompt Learning for Zero-shot 3D Anomaly Detection