Multi-view anomaly detection is crucial for industrial inspection to identify defects hidden from single viewpoints. Yet, despite their strong success in single-view applications, reconstruction-based methods remain largely underexplored in multi-view settings. A key challenge in extending these methods lies in generating view-consistent pseudo-anomalies. To address this, we propose PAR, a reconstruction-based framework for enhanced multi-view anomaly detection. First, we introduce a feature perturbation strategy coupled with a consensus aggregation module that integrates these perturbed features across viewpoints to mitigate view-inconsistent anomalies and construct highly aligned pseudo-anomaly representations. Second, we employ a view-specialized multi-decoder architecture; by assigning a dedicated decoder to each view, we increase the model's representational complexity, enabling the generation of highly feasible, localized anomaly maps rather than less precise, view-agnostic reconstructions. Last, we incorporate Fourier attention into the decoders to emphasize global frequency-domain structures over local token evidence, effectively preventing identity mapping. Extensive experiments on the Real-IAD dataset demonstrate that PAR consistently outperforms state-of-the-art methods in sample-level, image-level, and pixel-level anomaly detection. Our code is available at https://github.com/SpatialAILab/PAR_mvad.
Figures & tables
Figure 1. Overview of the key components of PAR. (a) Given multi-view inputs, feature perturbation first injects anomaly corruption. The consensus aggregation module then integrates these complementary cues across viewpoints to construct multi-view-consistent pseudo-anomaly features. (b) A shared decoder can produce view-agnostic reconstructions and inaccurate localization (red circles), whereas view-specialized decoders enable more accurate and consistent anomaly localization across viewpoints.
Figure 2. Overview of the proposed framework for multi-view anomaly detection. Multi-view images are first encoded by a pre-trained encoder. The resulting features are then perturbed by the composite feature perturbation (CFP) module using feature jitter, masking, and dropout to generate anomaly-like corrupted features. These perturbed features are aggregated across views by the consensus aggregation module to construct more view-aligned pseudo-anomaly features. Finally, the aggregated features are fed into a set of view-specialized decoders, where each decoder is responsible for reconstructing the normal feature of a particular view. Fourier attention is employed in each decoder to emphasize global structural patterns during reconstruction. The model is trained with the reconstruction loss Lrec .
Figure 3. A t -SNE visualization of latent pseudo-anomaly features generated by the vanilla individual-view perturbation strategy (red x markers) and our consensus aggregation method (blue dots). More compact feature distributions indicate pseudo-anomaly features are better aligned across views.
Figure 4. Comparison of singular values and weight distributions between the baseline single decoder and our proposed multi-view decoder. Higher singular values indicate greater representational capacity, while higher variance in the weight distribution suggests that the decoder learns more expressive and diverse parameters instead of collapsing around zero.
Figure 5. Visualization of reconstructed feature maps for different attention mechanisms. Successful normal-feature reconstruction should suppress the injected noise patch.
Class
Representation-based methods
Reconstruction-based methods
PatchCore
CFlow
FiCo
VSAD
Multi-Flow
DRAEM
DeSTSeg
RD++
CKAAD
RealNet
MVAD
PAR (ours)
Audiojack
93.7 /81.4/97.6
86.5/81.1/89.9
92.3/89.5/98.3
90.3/89.7/98.7
96.6 / 90.7 /97.7
91.4/83.2/92.8
88.5/81.9/95.6
86.4/85.4/86.3
87.8/89.2/95.3
91.3/79.8/91.4
93.1/86.9/ 98.8
93.3/ 91.9 / 99.5
Bottle cap
94.1/91.7/94.2
98.9 /86.8/95.3
98.8 / 97.7 /98.6
98.2/ 97.5 /98.6
98.0/94.6/99.1
96.3/67.6/88.1
96.6/85.3/95.1
96.4/95.0/98.5
95.1/92.8/97.6
97.3/92.8/98.5
98.9 /95.6/ 99.7
97.8/96.9/ 99.8
Button battery
81.3/79.2/96.7
93.3/77.8/92.0
81.9/77.4/91.2
96.4 / 93.8 /97.6
96.6 /89.4/98.6
94.6/84.4/93.1
92.0/90.8/97.0
94.4/89.0/95.9
91.1/85.7/96.8
90.6/82.5/96.6
93.4/90.8/ 98.8
94.7/ 93.1 / 99.3
End cap
86.7/80.6/96.2
83.2/75.2/89.8
89.0/82.7/95.9
94.7 /89.3/96.8
90.9/84.8/95.5
75.4/64.8/86.3
87.2/81.5/94.0
93.7/86.7/96.0
94.8 / 90.0 /96.4
81.4/72.8/92.0
91.9/85.8/ 98.1
94.4/ 90.5 / 99.4
Eraser
90.6/90.7/98.3
92.1/90.3/98.9
89.4/90.1/97.2
93.8/ 95.0 /99.1
92.8/92.1/98.8
69.4/70.4/81.3
95.6 / 96.3 /98.7
93.6/90.4/96.2
92.3/91.9/96.3
88.9/87.2/98.2
91.8/91.2/ 99.2
96.3 /94.4/ 99.7
Table 1. Comparison of quantitative experimental results on the Real-IAD dataset. Best value in each row for each metric is boldfaced, and second best is underlined. Scores are reported as S-AUROC / I-AUROC / P-AUROC.
Figure 6. Visualization of anomaly detection results achieved by SOTA models on the Real-IAD dataset.
Proposed methods
Real-IAD
MANTA
M 2 AD
Baseline
78.92 / 74.08 / 95.15
63.65 / 60.95 / 79.77
67.53 / 61.63 / 87.64
+ Consensus aggregation
96.20 / 91.94 / 99.08
96.58 / 89.01 / 96.88
91.31 / 85.48 / 95.52
+ View-specialized decoders
96.63 / 92.44 / 99.10
96.69 / 89.52 / 97.14
91.44 / 87.68 / 95.32
+ Fourier attention
96.73 / 92.62 / 99.12
96.78 / 89.60 / 97.15
92.84 / 89.18 / 95.90
Table 2. Component ablation study on three benchmark datasets. Metrics are reported as S-AUROC / I-AUROC / P-AUROC (%). The baseline uses feature perturbation without consensus aggregation, a single shared decoder, and linear attention. Consensus aggregation, view-specialized decoding, and Fourier attention are then added incrementally.
Random masking
Feature jitter
Dropout
S-AUROC (%)
I-AUROC (%)
P-AUROC (%)
95.71
91.98
98.97
✓
94.98
91.46
98.26
✓
95.89
92.00
98.91
✓
96.54
92.36
99.11
✓
✓
96.01
92.37
98.89
✓
✓
96.59
92.54
99.01
Table 3. Ablation study of combinations of pseudo-anomaly feature generation methods.
Softmax Attn.
Masked Attn.
Linear Attn.
Fourier Attn.
91.91 / 89.41 / 98.65
96.24 / 91.35 / 98.73
96.63 / 92.44 / 99.10
96.73 / 92.62 / 99.12
Table 4. Comparison with other attention mechanisms on Real-IAD. “Attn.” denotes attention.
Model
Hidden Dim.
FFN Depth
S-AUROC (%)
I-AUROC (%)
P-AUROC (%)
Single-decoder
3072
1
96.13
91.82
99.05
6144
2
95.57
91.82
99.02
Multi-decoder (ours)
3072
1
96.73
92.62
99.12
Table 5. Comparison between the single-decoder baseline and the proposed multi-decoder model. ‘Dim.’ denotes dimension, ‘FFN’ denotes feed-forward network.
Method
Number of missing views
0
1
2
3
4
Multi-Flow
95.85/90.27/96.40
57.01/54.00/79.86
56.90/52.03/71.48
55.67/48.91/62.26
54.96/48.57/56.31
MVAD
94.33/89.56/98.58
92.65 /89.33/98.52
91.36/89.06/98.49
88.93/88.28/98.29
88.92/88.93/98.48
Ours
96.73 / 92.62 / 99.12
92.45/ 89.96 / 98.84
91.56 / 89.97 / 98.85
90.35 / 89.56 / 98.84
91.71 / 91.71 / 98.85
Table 6. Performance comparison of multi-view anomaly detection methods under different numbers of missing views. Scores are reported as S-AUROC / I-AUROC / P-AUROC.
Pre-trained backbone
Type
S-AUROC (%)
I-AUROC (%)
P-AUROC (%)
D-iGPT ( Ren et al., 2023 )
MIM
96.58
92.16
98.54
MoCo v3 ( Chen et al., 2021 )
CL
96.02
91.56
98.50
iBOT ( Zhou et al., 2021 )
CL + MIM
96.49
92.64
98.83
DINOv2 ( Oquab et al., 2024 )
CL + MIM
96.73
92.62
99.12
Table A1 . Comparison across different pre-trained backbones. "MIM" denotes masked image modeling, and "CL" denotes contrastive learning.
Aggregation strategy
S-AUROC (%)
I-AUROC (%)
P-AUROC (%)
MLP-Mixer ( Tolstikhin et al., 2021 )
96.45
92.08
98.94
Cross-attention ( Vaswani et al., 2017 )
96.18
92.24
98.96
MLP ( Rumelhart et al., 1986 )
96.73
92.62
99.12
Table A2 . Ablation study on aggregation strategies.
View 1–View 2
View 1–View 3
View 1–View 4
View 1–View 5
Mean
w/o CA
0.1015
0.0419
0.0997
0.1122
0.0888
w/ CA
0.2328
0.1226
0.2575
0.2167
0.2074
Table A3 . Mutual information scores based on cross-covariance. “CA” denotes consensus aggregation.
View 1–View 2
View 1–View 3
View 1–View 4
View 1–View 5
w/o CA
9.22
2.32
7.36
2.94
w/ CA
31.22
11.99
37.35
15.79
Table A4 . Cross-view token retrieval accuracy (%) using View 1 as the anchor. “CA” denotes consensus aggregation.
Missing View 1
Missing View 2
Missing View 3
Missing View 4
Missing View 5
Ours
96.5 / 92.6 / 99.0
95.7 / 92.3 / 99.0
95.6 / 92.2 / 98.9
95.0 / 92.1 / 98.9
95.6 / 92.1 / 98.9
Table A5 . Performance under missing-view settings. Scores are reported as S-AUROC / I-AUROC / P-AUROC (%).
No transformation
Affine transformation
MVAD
90.08 / 86.42 / 97.71
63.24 / 63.67 / 90.41
Ours
96.73 / 92.62 / 99.12
89.59 / 89.73 / 98.03
Table A6 . Comparison under affine transformations. Scores are reported as S-AUROC / I-AUROC / P-AUROC (%).
GFLOPs
GPU Memory (MB)
FPS
Multi-Flow
36.896
315.97
369.58
MVAD
113.828
340.41
283.14
Ours
1085.83
3554.24
102.41
Table A7 . Model efficiency analysis.
Category
CDO ( Cao et al., 2023 )
RD++ ( Tien et al., 2023a )
MSFlow ( Zhou et al., 2024 )
INP-Former ( Luo et al., 2025 )
RealNet ( Zhang et al., 2024 )
CKAAD ( Fang et al., 2025 )
MVAD ( He et al., 2026 )
Ours
Bird
73.8/71.8/–
90.8 /71.3/70.2
86.8/61.8/62.0
87.7/71.8/67.2
90.6 / 76.3 / 97.1
77.6/ 76.3 /96.3
82.2/71.8/97.0
85.6/ 81.6 / 97.6
Car
84.1/75.7/–
86.3 /68.6/68.2
71.5/69.7/55.9
85.6/ 80.6 /53.9
79.0/65.2/ 92.3
77.6/75.4/89.8
84.5/77.8/91.9
90.3 / 89.0 / 93.1
Cube
95.1 / 91.1 /–
83.2/76.0/74.6
80.1/74.6/57.8
89.4 / 86.5 /74.5
94.5 /81.9/90.1
81.1/74.5/90.9
82.2/75.4/ 91.7
88.7/80.7/ 94.5
Dice
76.0/71.8/–
98.5 / 90.0 /89.4
79.1/77.5/69.4
72.5/74.4/83.7
97.3/88.6/ 94.0
98.4 /86.2/92.6
97.7/88.0/92.6
96.5/ 90.7 / 96.2
Doll
99.7 / 90.4 /–
67.4/66.0/65.9
57.3/56.2/55.1
99.0 / 84.4 /73.7
79.3/70.0/95.3
77.0/67.2/96.4
76.2/69.9/ 97.2
82.2/73.5/ 97.8
Holder
96.5/91.7/–
99.1 /87.8/87.8
97.9/68.6/76.6
80.5/70.0/76.4
98.3/84.6/91.7
97.2/ 94.4 /92.3
98.2/90.2/ 93.0
99.4 / 97.2 / 96.4
Table A8 . Comparison on the M 2 AD dataset. The best score for each metric in each row is boldfaced, and the second-best score is underlined. Scores are reported as S-AUROC / I-AUROC / P-AUROC. P-AUROC has not been reported for the CDO model ( Cao et al., 2023 ) .
Class
Embedding-based methods
Reconstruction-based methods
PatchCore
CFlow
FiCo
VSAD
Multi-Flow
DRAEM
DeSTSeg
RD++
CKAAD
RealNet
MVAD
Ours
Block Inductor
92.1/ 92.5 /97.8
84.3/87.2/95.7
77.7/69.2/87.5
93.2 / 93.7 / 98.6
89.9/82.3/95.1
85.6/81.8/71.8
92.1/91.4/97.7
91.0/91.4/96.3
89.5/91.2/97.2
76.1/77.4/88.6
90.3/92.0/98.3
95.5 /90.4/ 98.8
Copper Standoff
97.1/97.5/97.9
98.0/94.5/96.8
66.2/83.7/97.5
96.3/ 97.9 /98.1
97.0/91.7/92.8
66.9/61.7/71.7
91.8/92.8/94.7
98.5/ 98.0 /98.4
98.2/ 97.9 /98.1
92.9/87.1/95.3
99.2 /96.6/ 98.6
99.8 /96.4/ 99.2
Flat Nut
95.5/95.6/98.2
84.8/88.7/97.3
83.0/86.2/95.9
96.4 / 97.7 / 99.6
86.4/77.9/98.2
60.1/62.5/69.1
84.7/83.2/96.2
95.0/ 96.4 /98.8
89.8/92.6/96.2
83.8/81.4/94.5
90.5/92.9/99.2
98.3 /89.1/ 99.4
Led
96.9/97.8/98.6
97.9/97.8/99.1
98.8/98.2/98.6
99.0 / 99.8 /99.1
92.3/83.2/96.1
92.0/90.1/86.9
96.0/92.0/97.5
98.9/ 99.0 / 99.4
97.9/97.2/98.5
90.8/92.6/97.3
98.6/98.9/99.3
99.9 /92.8/ 99.6
Led Pad
98.7/97.1/97.7
91.4/94.0/97.1
98.8/96.7/98.5
95.0/95.4/97.2
89.9/81.1/94.8
60.7/58.8/71.7
98.3/97.0/95.2
96.4/97.1/96.9
98.9/ 97.2 /98.8
77.1/78.3/91.8
99.3 / 99.2 / 98.9
99.6 /89.5/ 99.2
Table A9 . Comparison on the MANTA dataset. The best score for each metric in each row is boldfaced, and the second-best score is underlined. Scores are reported as S-AUROC / I-AUROC / P-AUROC. The starred names abbreviate Round Button Cap, Square Button Cap, and Yellow Green Washer, respectively.
Figure A1 . Visualization of anomaly detection results on the Real-IAD dataset.
In multi-view anomaly detection, more cross-view information can actually hurt. When multiple inspection views are naively fused in a reconstruction-based pipeline, normal cues from intact views propagate to the decoder, which faithfully reconstructs anomalous regions, collapsing the reconstruction gap the detector depends on. We call this failure mode cross-view information leakage and show that effective multi-view fusion must explicitly restrict the information reaching the decoder. Building on this insight, we present GLAD(Global-Local Attention Driven framework), the first framework combining vision foundation model features with local and global cross-view fusion for multi-view anomaly detection. The Multi-view Merging Attention (MMA) module performs local cross-view fusion at linear complexity with learnable view importance weighting and token-wise gating, letting each view selectively incorporate fine-grained evidence from other views at O(N) cost. The Object-Guided Attention (OGA) module captures global context by aggregating class tokens from all views into a single object-level representation and broadcasting it back to patch tokens via temperature-scaled sigmoid gating, replacing the original patch representations rather than adding a residual to preserve the reconstruction gap. Experiments on Real-IAD and MANTA-Tiny show that GLAD outperforms state-of-the-art methods across sample-, image-, and pixel-level metrics, confirming that principled information restriction is key to multi-view anomaly reasoning.
Shang-Fu Chen, Kuan-Chuan Peng, Jhih-Ciang Wu +2
National Taiwan University Taipei, Taiwan · Mitsubishi Electric Research Laboratories (MERL) Cambridge, MA, USA · National Taiwan Normal University Taipei, Taiwan +2
Existing Multi-view Anomaly Detection (MAD) methods assume that all views are completely available and model each view separately. However, in real industrial scenarios, information in the view may be missing due to faults such as occlusion, which leads to the performance degradation of existing methods due to the lack of a multi-view consistency prior. To address this, we explored a more challenging task: Incomplete Multi-View Anomaly Detection (IMVAD), in which some areas of each view were masked. We proposed a pipeline for automatically generating the IMVAD dataset and generated the RIMAD dataset based on the Real-IAD dataset through this pipeline. In addition, in order to effectively utilize the information of multiple views in the absence of view information, we propose IMMoE, which consists of two key modules: (1) Multi-View Expert Fusion (MVEF) effectively fuses multi-view information through a multi-view expert network and guides the reconstruction of a single view; (2) Local Anomaly Enhancement Encoder (LAEE) effectively prevents the model from overfitting the mask region by applying dropout to local features. Our method achieves state-of-the-art performance on both the RIMAD and Real-IAD datasets, especially on RIMAD, we have increased the pixel-level and image-level metrics by 11.8% and 2.8%, respectively. Our source code is available at https://github.com/HULEI7/IMMoE
Lei Hu
South China University of Technology, Guangzhou, China
Pose-agnostic Anomaly Detection (PAD) remains challenging as anomalies can appear under arbitrary viewpoints, requiring methods to handle significant pose variations. Existing approaches rely on complex 3D reconstruction, which are computationally expensive and require extensive multi-view data. We propose PADFormer, a novel image-space approach that leverages Vision Transformer (ViT) to directly reconstruct anomaly-free versions of query images while preserving pose information. Our key insight is to adapt cross-view masked reconstruction for anomaly detection through training exclusively on normal data, combined with dynamic patch selection and spatial alignment mechanisms that enable effective learning from sparse reference views under significant pose variations. During inference, we perform multiple forward passes with different masking patterns to generate an ensemble of anomaly-free reconstructions, ensuring comprehensive coverage of the query image. Anomalies are detected by comparing these reconstructions with the query image. PADFormer achieves state-of-the-art results on the PAD benchmark while maintaining comparable performance on classic few-shot anomaly detection (FSAD) tasks, demonstrating superior efficiency and generalization without requiring 3D reconstruction.
Ruiqi Wang, Yiming Qian, Fenggen Yu +4
Amazon · Simon Fraser University · ⋆Work carried out during internship at Amazon. +1