The quest for incremental unified multimodal anomaly detection seeks to empower a single model with the ability to systematically detect anomalies across all categories and support incremental learning to accommodate emerging objects/categories. Central to this pursuit is resolving the catastrophic forgetting dilemma, which involves acquiring new knowledge while preserving prior learned knowledge. Despite some efforts to address this dilemma, a key oversight persists: ignoring the potential impact of spurious and redundant features on catastrophic forgetting. In this paper, we delve into the negative effect of spurious and redundant features on this dilemma in incremental unified frameworks, and reveal that under similar conditions, the multimodal framework developed by naive aggregation of unimodal architectures is more prone to forgetting. To address this issue, we introduce a novel denoising framework called IB-IUMAD, which exploits the complementary benefits of the Mamba decoder and information bottleneck fusion module: the former dedicated to disentangle inter-object feature coupling, preventing spurious feature interference between objects; the latter serves to filter out redundant features from the fused features, thus explicitly preserving discriminative information. A series of theoretical analyses and experiments on MVTec 3D-AD and Eyecandies datasets demonstrates the effectiveness and competitive performance of IB-IUMAD.
Figures & tables
Figure 1: Top shows the performance of IUMAD task on the MVTec 3D-AD dataset, where catastrophic forgetting occurs significantly. Bottom is the comparison of our incremental unified multimodal framework with previous paradigms.
Figure 2: The impact of spurious and redundant features on catastrophic forgetting in incremental unified frameworks.
Figure 3: The overall framework of IB-IUMAD, where the Mamba decoder and information bottleneck fusion module (IBFM) are the core designs of this work, aims to mitigate the effects of spurious and redundant features on catastrophic forgetting.
10-0 with 0 step (setting 1)
9-1 with 1 step (setting 2)
6-4 with 1 step (setting 3)
6-1 with 4 steps (setting 4)
Method
I-AUROC
AUPRO
FM ( ⇓ )
I-AUROC
AUPRO
FM ( ⇓ )
I-AUROC
AUPRO
FM ( ⇓ )
I-AUROC
AUPRO
FM ( ⇓ )
RGB
IUF (ECCV24)
87.1
88.5
– / –
82.4
85.1
2.9 / 1.5
76.7
78.6
8.3 / 6.9
73.9
76.3
12.5 / 7.9
CDAD (CVPR25)
76.4
87.5
– / –
74.2
83.9
2.1 / 1.9
68.9
77.5
7.4 / 6.3
67.4
73.9
8.0 / 6.7
IB-IUMAD
88.9
89.3
– / –
85.3
87.6
1.8 / 1.3
79.8
81.2
6.5 / 5.8
75.7
80.5
8.7 / 6.3
3D
IUF (ECCV24)
73.7
65.4
– / –
70.4
62.1
1.6 / 1.4
67.4
59.2
4.4 / 3.6
66.8
59.7
4.7 / 3.0
CDAD (CVPR25)
60.6
40.6
– / –
58.8
39.2
1.1 / 0.2
55.5
37.9
1.7 / 0.9
56.1
38.5
2.5 / 1.4
Table 1: I-AUROC/AUPRO results of our approach on MVTec 3D-AD under four incremental unified anomaly detection settings.
10-0 with 0 step (setting 1)
9-1 with 1 step (setting 2)
6-4 with 1 step (setting 3)
6-1 with 4 steps (setting 4)
Method
I-AUROC
AUPRO
FM ( ⇓ )
I-AUROC
AUPRO
FM ( ⇓ )
I-AUROC
AUPRO
FM ( ⇓ )
I-AUROC
AUPRO
FM ( ⇓ )
RGB
IUF (ECCV24)
77.2
84.0
– / –
73.9
79.6
3.8 / 3.5
62.9
68.6
12.4 / 9.1
63.6
69.1
10.6 / 8.8
CDAD (CVPR25)
78.1
82.1
– / –
73.2
76.5
4.2 / 3.1
63.7
70.4
9.5 / 7.7
65.5
72.3
8.4 / 6.3
IB-IUMAD
78.5
85.3
– / –
74.1
83.2
2.7 / 1.1
65.4
71.2
8.1 / 6.5
66.5
72.7
7.3 / 5.7
3D
IUF (ECCV24)
57.2
39.4
– / –
55.6
36.4
2.1 / 0.8
51.5
37.4
3.8 / 1.2
51.9
38.2
3.0 / 1.1
CDAD (CVPR25)
58.4
30.2
– / –
57.1
29.3
0.9 / 0.1
53.1
29.6
4.2 / -0.4
52.9
30.1
4.4 / 0.3
Table 2: I-AUROC/AUPRO results of our approach on Eyecandies under four incremental unified anomaly detection settings.
Figure 6Figure 7
Ablation
6-1 with 4
6-4 with 1
9-1 with 1
10-0 with 0
Mamba
IBFM
I-AUROC
FM
I-AUROC
FM
I-AUROC
FM
I-AUROC
✗
✗
75.3
12.7
79.9
8.7
85.1
4.6
86.7
✓
✗
76.0
11.0
80.8
7.6
86.0
4.0
88.8
✗
✓
76.9
10.2
81.5
7.1
86.8
3.2
89.2
✓
✓
78.6
9.3
82.4
6.8
87.5
2.3
91.0
Table 5: The ablation study results on the MVTec 3D-AD dataset.
MVTec 3D-AD
Fusion types
I-AUROC
P-AUROC
AUPRO
Addition
89.8
96.7
88.6
ConcatFC
88.2
96.4
89.5
LinearGLU
86.7
95.8
88.3
Cross-attention
91.0
97.6
90.4
Table 6: The impact of multimodal fusion operations.
MVTec 3D-AD
Method
Frame Rate
Memory
I-AUROC
BTF
3.197
3810.6
86.5
AST
4.966
4639.4
93.7
M3DM
0.514
65261.2
94.5
IB-IUMAD
21.427
1483.7
91.0
Table 7: Comparison results in terms of the computing efficiency.
Current advancements in Multimodal Anomaly Detection (MAD) are largely driven by enhancing multimodal fusion, particularly through the integration of RGB and Depth data for richer anomaly representation. However, less attention was devoted to analyzing the role of cross-modal fusion bias, a well-known challenge in multimodal learning, in MAD. This gap motivates a key question: can we overcome this bias to break the performance bottleneck of current work? In this paper, we first analyze the impact of cross-modal fusion bias in MAD via the Fisher Information Matrix. Then, grounded in these findings, we propose UCFB, a simple yet effective plug-and-play framework designed to mitigate cross-modal fusion bias in MAD. It achieves this by jointly employing Fisher-information-guided dynamic calibration to adjust modality-specific regularization weights and canonical similarity analysis to improve inter-modal interactions. Extensive experiments on the MVTec 3D-AD and Eyecandies datasets demonstrate that UCFB achieves consistent improvements in single-class, multi-class, and few-shot settings.
Kaifang Long, Lianbo Ma, Liming Liu +1
Northeastern University Shenyang, China · CATL Ningde, China
Unsupervised Continuous Anomaly Detection (UCAD) is gaining attention for effectively addressing the catastrophic forgetting and heavy computational burden issues in traditional Unsupervised Anomaly Detection (UAD). However, existing UCAD approaches that rely solely on visual information are insufficient to capture the manifold of normality in complex scenes, thereby impeding further gains in anomaly detection accuracy. To overcome this limitation, we propose an unsupervised continual anomaly detection framework grounded in multimodal prompting. Specifically, we introduce a Continual Multimodal Prompt Memory Bank (CMPMB) that progressively distills and retains prototypical normal patterns from both visual and textual domains across consecutive tasks, yielding a richer representation of normality. Furthermore, we devise a Defect-Semantic-Guided Adaptive Fusion Mechanism (DSG-AFM) that integrates an Adaptive Normalization Module (ANM) with a Dynamic Fusion Strategy (DFS) to jointly enhance detection accuracy and adversarial robustness. Benchmark experiments on MVTec AD and VisA datasets show that our approach achieves state-of-the-art (SOTA) performance on image-level AUROC and pixel-level AUPR metrics.
Mingle Zhou, Jiahui Liu, Jin Wan +2
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 · Faculty of Data Science, City University of Macau, Macau, China
Multi-modal industrial anomaly detection typically relies on separate models for each product category, fundamentally limiting practical scalability. When shifting to a unified paradigm that handles diverse classes simultaneously, detection accuracy often degrades due to inter-class interference and feature manifold confusion. To overcome these challenges, we propose a Unified Reference guided Cross-modal Mapping framework, named Uni-RCM. At its core, we propose a reference guide block to dynamically filter out category-specific noise by introducing a learnable reference feature, which captures the commonalities across different modalities. Besides, an offline residual quantizer is proposed to characterize the normal distribution by multiple cascaded codebooks. Extensive evaluations on the MVTec-3D AD dataset demonstrate the state-of-the-art performance in the challenging multi-class setting and in terms of image-level detection and pixel-level localization.
Yangchen Wu, Huiqiang Xie
School of Information Science and Technology, Jinan University, Guangzhou, Guangdong 510632, P. R. China