Source-linked AI summary
See More, Detect Less? Taming Information Leakage in Multi-View Anomaly Detection
Shang-Fu Chen, Kuan-Chuan Peng, Jhih-Ciang Wu, Wen-Huang Cheng, Kai-Lung Hua
TL;DR
Reconstruction-based multi-view anomaly detection can suffer from cross-view information leakage, as normal cues from intact views may help reconstruct anomalous regions and collapse the reconstruction gap. GLAD combines vision foundation-model features with restricted local and global fusion, and experiments on Real-IAD and MANTA-Tiny show improvements over state-of-the-art methods across sample-, image-, and pixel-level metrics.
Problem
Naive cross-view fusion can propagate normal cues to the decoder, collapsing the reconstruction gap needed to detect anomalies.
Method
GLAD uses MMA for linear-complexity local fusion and OGA for globally conditioned, information-restricted patch representations.
Results
GLAD outperforms state-of-the-art methods on Real-IAD and MANTA-Tiny across sample-, image-, and pixel-level metrics.
Takeaways & Limitations
The paper identifies explicit information restriction as essential for preserving the reconstruction gap in robust multi-view anomaly reasoning.
Abstract
from arXiv · showhide
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 \emph{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 $\mathcal{O}(N)$ 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.
1 Introduction
GLAD addresses the failure of naive cross-view fusion in reconstruction-based anomaly detection by combining foundation-model features with restricted global-local fusion. It targets efficient inter-view reasoning while preserving the reconstruction gap and reports improvements across multiple evaluation levels.
- Motivation: Industrial inspection uses multiple camera angles for surface coverage, but many anomaly-detection methods treat views independently and ignore cross-view relationships.Multi-view anomaly detection is therefore important for industrial quality control.
- Framework: GLAD combines shared frozen vision-foundation-model encoders with global and local cross-view fusion before decoding.The framework preserves single-view representations while enabling structured inter-view interactions.
- Framework: MMA performs local fusion at O(N) rather than O(N^2) complexity using learnable view-importance weighting and adaptive token-wise gating.This design addresses the computational cost of naive cross-attention over tokens and views.
- Problem: Cross-view information leakage occurs when normal cues from intact views let the decoder reconstruct anomalous regions, collapsing the normal–abnormal reconstruction gap.GLAD instead uses information-restricted fusion to restore the gap.
- Framework: OGA restricts decoder access by replacing original patch representations with globally conditioned outputs instead of adding residual features.Temperature-scaled sigmoid gating further limits diffuse information spread and avoids competition among normal tokens.
- Results: GLAD improves consistently over single-view baselines and state-of-the-art methods at sample-, image-, and pixel-level metrics on MANTA and Real-IAD.The supplied introduction reports the cross-level comparison without specifying aggregate values.
2 Related Work
Prior anomaly-detection methods differ in modeling scope, backbone choice, and anomaly-detection strategy, while related class-token methods primarily address single-image settings. GLAD is positioned against these categories as a multi-view foundation-model approach.
- Anomaly Detection: Single-view anomaly-detection methods are grouped into class-specific models, unified conventional-backbone models, and unified vision-foundation-model models.The categories are labeled G0, G1, and G2, respectively.
- Anomaly Detection: Unified conventional-backbone methods reduce deployment overhead but struggle with diverse appearances in multi-view settings.The passage contrasts their scalability with limitations of conventional backbones.
- Anomaly Detection: Class-specific methods can achieve competitive accuracy but incur scalability costs, while synthesis-based approaches face a synthetic-to-real gap across viewpoints.Embedding-based methods also lack cross-view spatial correspondence.
- Anomaly Detection: Reconstruction-based approaches reproduce normal patterns and flag deviations, but without cross-view constraints may miss anomalies visible from only certain viewpoints.This limitation motivates explicit multi-view modeling.
- Related Attention Methods: Class-token attention and learned query-token methods provide related fusion patterns, but the cited methods operate on single images rather than multi-view anomaly detection.GLAD extends the broader design space toward multi-view reasoning.
3 Our Proposed Method – GLAD
GLAD inserts restricted global and local cross-view fusion between frozen per-view encoding and reconstruction, using OGA for object context and MMA for fine-grained evidence. Its design replaces or selectively gates information to preserve anomaly-sensitive reconstruction differences while enabling efficient multi-view reasoning.
- 3.1 Overview: GLAD independently encodes each view with a shared frozen Vision Transformer, then fuses token sequences before a bottleneck and Transformer decoder reconstructs features.The encoder provides class tokens for global content and patch tokens for local spatial information; anomaly maps use encoder–decoder cosine distances.
- 3.2 Object-Guided Attention (OGA): OGA concatenates class tokens across views into one learned object-level representation, cross-attends it over all patch tokens, and broadcasts global context back to views.The resulting representation is designed to capture semantics shared across viewpoints before token-level fusion.
- 3.2 Object-Guided Attention (OGA): OGA replaces original patch tokens with cross-attention outputs rather than adding residuals, preventing normal templates from diluting anomaly signals.This token-replacement mode explicitly restricts information reaching the reconstruction decoder and preserves the reconstruction gap.
- 3.2 Object-Guided Attention (OGA): OGA uses independently applied temperature-scaled sigmoid gates, allowing abnormal patches to receive distinct activation levels without competition from the normal majority.The temperature is set to τ = 0.7, and sigmoid gating avoids softmax normalization across patches.
- 3.3 Multi-view Merging Attention (MMA): MMA enriches tokens with fine-grained cross-view evidence, strengthening reconstruction gaps for anomalies and supporting localization and detection.Queries come from each view’s tokens, while keys and values are aggregated from the remaining views using learned importance scores.
- 3.3 Multi-view Merging Attention (MMA): MMA performs local cross-view fusion with linear attention, learned view-importance weighting, and token-wise gating so each view selectively incorporates evidence from other views.It uses a ReLU2 kernel and avoids materializing the full V N × V N attention matrix, reducing the attention computation to linear dependence on token count.
4 Experiments
GLAD is evaluated on Real-IAD and MANTA-Tiny using multi-granularity metrics and ablations of its fusion and information-restriction mechanisms. It consistently improves detection and localization while maintaining efficiency and robustness to missing views.
- Experimental Setup: GLAD uses a frozen DINOv2 ViT-Base/16 encoder and Transformer decoder, with MMA and OGA providing local and global cross-view fusion.Intermediate features from layers 2 through 9 are averaged into one token sequence per view.
- Metrics: The evaluation reports sample-, image-, and pixel-level metrics, including AUROC, AP, F1, and P-AUPRO for defective-region localization.Sample-level scores aggregate evidence across views, whereas image-level scores evaluate views independently.
- Comparison with the State-of-the-Art: GLAD achieves the best results across all 10 MANTA-Tiny metrics, including pixel-level scores of 95.2/47.6/49.6/85.9.Compared with Dinomaly, MANTA-Tiny P-AP improves by +2.8 (44.8→47.6) and P-F1 by +1.3 (48.3→49.6).
- Qualitative Results: GLAD produces cleaner anomaly maps with fewer false positives and more focused activation on ground-truth defect regions.The qualitative comparison covers Real-IAD and MANTA, with warmer colors indicating higher anomaly scores.
- Ablation Study: MMA improves sample-level detection, while OGA further boosts pixel-level localization in the module ablation.On Real-IAD, MMA alone improves S-AUROC by +2.2, while adding OGA raises P-AP from 42.7 to 44.7 (+2.0).
- Ablation Study: 2.66M parameters and 12.04 ms per forward pass make linear attention the most efficient MMA fusion alternative.It uses substantially less training memory than MVAS while achieving nearly identical pixel-level accuracy.
- Ablation Study: Information-restriction variants improve MANTA-Tiny while keeping sample-level metrics stable or better, and GLAD remains largely stable with fewer views.The restrictions include token replacement, sigmoid gating, withheld global-token feedback, and sharpened temperature.
5 Conclusion
The paper presents GLAD as a global-local framework for multi-view anomaly detection and identifies cross-view information leakage as a central challenge. Experiments on Real-IAD and MANTA show that GLAD outperforms state-of-the-art methods, while information restriction preserves the reconstruction gap needed for detection.
- 5 Conclusion: GLAD combines global and local cross-view learning through OGA and MMA for multi-view anomaly detection.OGA restricts information to preserve the reconstruction gap, while MMA provides efficient token-level fusion with linear complexity.
- 5 Conclusion: Naive fusion can collapse the reconstruction gap by passing normal cues from intact views across views to the decoder.The paper names this failure mode cross-view information leakage.
- 5 Conclusion: GLAD outperforms state-of-the-art methods on the Real-IAD and MANTA benchmarks.The conclusion frames information restriction as a perspective for future multi-view reasoning in reconstruction-based settings beyond anomaly detection.