Multimodal Industrial Anomaly Detection via Geometric Prior

Min Li; Jinghui He; Gang Li; Jiachen Li; Jin Wan; Delong Han

cs.CVMar 24, 2026

Multimodal Industrial Anomaly Detection via Geometric Prior

Authors: Min Li, Jinghui He, Gang Li, Jiachen Li, Jin Wan, Delong Han

Organizations: Faculty of Data Science, City University of Macau, Macau, China. · Key Laboratory of Computing Power Network and Information Security, Ministry of Education, Shandong Computer Science Center (National Supercomputer Center in Jinan), Qilu University of Technology (Shandong Academy of Sciences), Jinan, China. · Shandong Provincial Key Laboratory of Computing Power Internet and Service Computing, Shandong Fundamental Research Center for Computer Science, Jinan, China.

Abstract

The purpose of multimodal industrial anomaly detection is to detect complex geometric shape defects such as subtle surface deformations and irregular contours that are difficult to detect in 2D-based methods. However, current multimodal industrial anomaly detection lacks the effective use of crucial geometric information like surface normal vectors and 3D shape topology, resulting in low detection accuracy. In this paper, we propose a novel Geometric Prior-based Anomaly Detection network (GPAD). Firstly, we propose a point cloud expert model to perform fine-grained geometric feature extraction, employing differential normal vector computation to enhance the geometric details of the extracted features and generate geometric prior. Secondly, we propose a two-stage fusion strategy to efficiently leverage the complementarity of multimodal data as well as the geometric prior inherent in 3D points. We further propose attention fusion and anomaly regions segmentation based on geometric prior, which enhance the model's ability to perceive geometric defects. Extensive experiments show that our multimodal industrial anomaly detection model outperforms the State-of-the-art (SOTA) methods in detection accuracy on both MVTec-3D AD and Eyecandies datasets.

Figures & tables

Fig. 1: Comparison of three MIAD pipelines: (a) early fusion based, (b) intermediate fusion based, and (c) geometric prior based. Our framework integrates a dual fusion strategy, using a 3D expert model to extract geometric prior, which are then used to guide multimodal fusion and abnormal region segmentation.

Fig. 2: Architecture of GPAD. RGB-D features are extracted and quantized through a pretrained encoder, capturing both color and texture information as features $\mathbf{\mathit{}}{F_{T}}$ , as well as high-level semantic features $\mathbf{\mathit{}}{F_{S}}$ . Point clouds and their normals are processed by a point cloud expert model to extract fine-grained geometric features $\mathbf{\mathit{}}{F_{p}}$ . These features are then fused with RGB-D features $\mathbf{\mathit{}}{F_{T}}$ and $\mathbf{\mathit{}}{F_{S}}$ via geometry-conditioned attention, fully leveraging the complementary information between modalities. The SRN module refines multimodal features, removing redundant information and enhancing anomaly detection through image reconstruction. $\hat{I}$ and $\hat{I}_{a}$ are the reconstructed images without and with anomalies, respectively. PCUnet employs geometric prior to guide the precise segmentation of anomaly regions.

Fig. 3: Geometry-Conditioned Cross-Modal Fusion. Each group of point cloud features is projected to the corresponding region in the image. The weighted point cloud features $F_{W}^{T}$ and $F_{W}^{S}$ are fused with $F^{{}^{\prime}}_{T}$ and $F_{S}$ , respectively.

Fig. 4: PCUnet architecture with geometric fusion. The geometric prior $F_{p}$ is adaptively projected and injected in each decoder layer to enhance the geometry awareness of the segmentation network.

	Method	Bagel	Cable Gland	Carrot	Cookie	Dowel	Foam	Peach	Potato	Rope	Tire	Mean
3D	VoxelAE [ 46 ]	69.3 / 26.0	42.5 / 34.1	51.5 / 58.1	79.0 / 35.1	49.4 / 50.2	55.8 / 23.4	53.7 / 35.1	48.4 / 65.8	63.9 / 1.5	58.3 / 18.5	57.1 / 34.8
	DepthGAN [ 46 ]	53.0 / 11.1	37.6 / 7.2	60.7 / 21.2	60.3 / 17.4	49.7 / 16.0	48.4 / 12.8	59.5 / 0.3	48.9 / 4.2	53.6 / 44.6	52.1 / 7.5	52.3 / 14.3
	DepthAE [ 46 ]	46.8 / 14.7	73.1 / 6.9	49.7 / 29.3	67.3 / 21.7	53.4 / 20.7	41.7 / 18.1	48.5 / 16.4	54.9 / 6.6	56.4 / 54.5	54.6 / 14.2	54.6 / 20.3
	FPFH [ 48 ]	82.5 / 97.3	55.1 / 87.9	95.2 / 98.2	79.7 / 90.6	88.3 / 89.2	58.2 / 73.5	75.8 / 97.7	88.9 / 98.2	92.9 / 95.6	65.3 / 96.1	78.2 / 92.4
	M3DM [ 4 ]	94.1 / 94.3	65.1 / 81.8	96.5 / 97.7	96.9 / 88.2	90.5 / 88.1	76.0 / 74.3	88.0 / 95.8	97.4 / 97.4	92.6 / 95.0	76.5 / 92.9	87.4 / 90.6
	Shape-Guided [ 3 ]	98.3 / 97.4	68.2 / 87.1	97.8 / 98.1	99.8 / 92.4	96.0 / 89.8	73.7 / 77.3	99.3 / 97.8	97.9 / 98.3	96.6 / 95.5	87.1 / 96.9	91.6 / 93.1

TABLE I: Comparison of GPAD and Other Models on the MVTec3D Dataset (I-AUROC% / AUPRO%).

Method	Bagel	Cable Gland	Carrot	Cookie	Dowel	Foam	Peach	Potato	Rope	Tire	Mean
PatchCore+FPFH [ 48 ]	99.6	99.2	99.7	99.4	98.1	97.4	99.6	99.8	99.4	99.5	99.2
M3DM [ 4 ]	99.5	99.3	99.7	98.5	98.5	98.4	99.6	99.4	99.7	99.6	99.2
Shape-Guided [ 3 ]	-	-	-	-	-	-	-	-	-	-	99.6
3DSR [ 2 ]	-	-	-	-	-	-	-	-	-	-	99.5
GPAD (Ours)	99.6	99.6	99.9	98.7	99.8	99.9	99.8	99.4	99.5	99.6	99.6

TABLE II: Comparison of GPAD and Other Models on the MVTec3D Dataset (P-AUROC%).

Fig. 5: Qualitative results for each class of the MVTec-3D AD [ 46 ] dataset. The color bar represents anomaly scores from 0 (normal, blue) to 1 (anomalous, red) using the jet colormap. Compared with 3DSR [ 2 ] , GPAD can obtain a more accurate segmentation region.

Category	M3DM [ 4 ]	3DSR [ 2 ]	CFM [ 35 ]	2M3DF [ 1 ]	GPAD (ours)
Can.C.	62.4	65.1	68.0	75.3	71.0
Cho.C.	95.8	99.8	93.1	95.5	100
Cho.P.	95.8	90.4	95.2	93.7	92.8
Conf.	100	97.8	88.0	96.7	96.0
Gum.B.	88.6	87.5	86.5	90.1	79.5
Haz.T.	75.8	86.1	78.2	79.2	84.8

TABLE III: Comparison of GPAD and Other Models on the Eyecandies Dataset (I-AUROC%).

Config.	Bagel	Cable Gland	Carrot	Cookie	Dowel	Foam	Peach	Potato	Rope	Tire	Mean
I-AUROC
w/o Norm	99.9	95.3	100	99.5	99.5	99.4	98.6	98.0	100	95.2	98.5
w/o GCA Fusion	99.7	95.0	99.9	98.3	99.6	99.0	99.5	98.3	100	96.7	98.6
w/o Geo Prior	99.7	95.0	99.7	99.5	99.7	99.6	98.8	98.7	99.7	97.1	98.7
w/o Depth	99.3	93.7	99.4	98.7	99.2	98.7	99.2	98.3	99.1	97.5	98.3
w/o Points	99.1	92.7	99.3	98.8	99.5	98.4	97.4	97.1.2	99.7	96.1	97.8

TABLE IV: Performance Comparison of Model Configurations.

Num	Bagel	Cable Gland	Carrot	Cookie	Dowel	Foam	Peach	Potato	Rope	Tire	Mean
I-AUROC
100	99.7	92.8	99.9	99.3	99.6	99.4	99.3	98.2	100	99.4	98.8
200	99.8	96.8	98.8	99.9	99.8	98.7	99.9	98.1	100	97.2	98.9
300	99.7	96.7	99.4	99.3	99.6	99.5	99.5	98.5	100	97.2	98.9
400	99.8	96.1	100	98.9	99.9	99.2	99.6	98.4	99.7	95.9	98.8
500	99.8	93.7	99.8	99.1	99.6	99.1	99.5	98.8	100	96.5	98.6

TABLE V: Comparison of Results for Different Group Quantities in Point Cloud Samples.

Fig. 6: Comparison of detection results (I-AUROC and AUPRO) on the MVTec-3D AD dataset when each sample is divided into 100, 200, 300, 400, and 500 groups.

Method	Memory Consumption	Inference	Mean
Method	(MBs)	(FPS)	I-AUROC(%)
BTF [ 51 ]	228.984	3.91	86.5
AST [ 52 ]	463	4.78	93.7
Shape-Guided [ 3 ]	237.58	2.1	94.7
M3DM [ 4 ]	6528.7	0.514	94.5
CPMF [ 7 ]	2195	0.609	95.1

TABLE VI: Inference speed and memory consumption comparison.

Explore similar work

Real-IAD MVN: A Multi-View Normal Vector Dataset and Benchmark for High-Fidelity Industrial Anomaly Detection
May 8, 2026Wenbing Zhu, Jianing Liang, Linjie Cheng +7Industrial Anomaly Detection Multimodal Anomaly Detection
1Fudan University · 4Rongcheer Co., Ltd. · 3Donghua University +1
CoGeoAD: Hierarchical Color-Geometric Fusion with Multi-View Attention for Zero-Shot 3D Anomaly Detection
Jun 24, 2026Ke Xu, Xinle Wang, Yanning Hou +3Industrial Anomaly Detection Multimodal Anomaly Detection
State Key Laboratory of Opto-Electronic Information Acquisition and Protection Technology, Anhui University, Hefei, China · School of Artificial Intelligence, Anhui University, Hefei, China · College of Intelligence Science and Technology, National University of Defense Technology, Changsha, China +1
Cross-Modal Mapping and Dual-Branch Reconstruction for 2D-3D Multimodal Industrial Anomaly Detection
Mar 4, 2026Radia Daci, Vito Renò, Cosimo Patruno +4Industrial Anomaly Detection Multimodal Anomaly Detection
CNR-ISASI, Institute of Applied Sciences and Intelligent Systems / National Research Council of Italy, via Monteroni snc, Lecce, 73100, Italy · CNR-STIIMA, Institute of Intelligent Industrial Technologies and Systems for Advanced Manufacturing, Via G. Amendola 122 D/O, Bari, 70126, Italy · Institute d’Electronique de Microelectronique et de Nanotechnologie (IEMN) / Universite Polytechnique Hauts de France; Universite de Lille; CNRS, UMR 8520, City, 59313, France

May 8, 2026J/K move · Enter open · S save

Real-IAD MVN: A Multi-View Normal Vector Dataset and Benchmark for High-Fidelity Industrial Anomaly Detection

Wenbing Zhu, Jianing Liang, Linjie Cheng, Yurui Pan+6

1Fudan University · 4Rongcheer Co., Ltd. · 3Donghua University+1

Industrial Anomaly Detection Multimodal Anomaly Detection Anomaly Detection

Open paper AskPDF arXiv

Industrial Anomaly Detection (IAD) is critical for quality control, but existing methods struggle with subtle, geometric defects. Standard 2D (RGB) images are sensitive to texture and lighting but often miss fine geometric anomalies. While 3D point clouds capture macro-shape, they are typically too sparse to detect micro-defects like scratches or pits. We address this fundamental data limitation by introducing Real-IAD-MVN (Multi-View Normal), a large-scale industrial dataset. By upgrading our acquisition system, Real-IAD-MVN captures high-fidelity surface normal maps from five distinct viewpoints, replacing sparse 3D data entirely. This provides a comprehensive geometric representation at a micro-detail level, making previously invisible side-wall and occluded defects explicitly detectable. Our experiments, conducted on this new dataset, first provide evidence that incorporating dense, multi-view pseudo-3D (surface normals) yields significantly better detection performance than using sparse 3D point cloud data. To further validate the dataset and provide a strong benchmark, we introduce a baseline method based on reconstruction, which learns to extract cross-modal unified prototypes from the image and normal map streams. We demonstrate that this unified prototype approach surpasses existing state-of-the-art multimodal fusion methods, highlighting the rich potential of our new dataset for advancing geometric anomaly detection.

Abstract

Explore similar work

Real-IAD MVN: A Multi-View Normal Vector Dataset and Benchmark for High-Fidelity Industrial Anomaly Detection

CoGeoAD: Hierarchical Color-Geometric Fusion with Multi-View Attention for Zero-Shot 3D Anomaly Detection

Cross-Modal Mapping and Dual-Branch Reconstruction for 2D-3D Multimodal Industrial Anomaly Detection