See More, Detect Less? Taming Information Leakage in Multi-View Anomaly Detection

Technical Summary: See More, Detect Less? Taming Information Leakage in Multi-View Anomaly Detection

Problem Statement
In industrial inspection, objects are often captured from multiple viewpoints to ensure complete surface coverage. While multi-view data offers complementary evidence, existing multi-view anomaly detection (AD) methods often fail to leverage this effectively. The paper identifies a critical failure mode termed cross-view information leakage. In standard reconstruction-based AD frameworks, a frozen Vision Transformer (ViT) encoder extracts features, and a lightweight decoder is trained to reconstruct normal patterns. Anomalies are detected via reconstruction errors. However, when multiple views are fused via simple residual addition, normal cues from intact views propagate to the decoder. This allows the decoder to "repair" anomalous regions using information from other views, faithfully reconstructing defects and collapsing the reconstruction gap essential for detection. Consequently, "seeing more" (more views) paradoxically leads to "detecting less."

Methodology: The GLAD Framework
The authors propose GLAD (Global-Local Attention Driven), a framework designed to integrate vision foundation model features with structured cross-view fusion while explicitly restricting information flow to prevent leakage. GLAD operates on an encoder-bottleneck-decoder paradigm using a frozen pre-trained ViT (DINOv2) and a trainable Transformer decoder. The core innovation lies in the fusion stage between the encoder and the bottleneck, comprising two complementary modules:

1. Object-Guided Attention (OGA) for Global Context:

   • Goal: To establish a holistic understanding of the object across all views before local fusion.
   • Mechanism: OGA aggregates class tokens ($z_{cls}$) from all $V$ views into a single global object token via a learned projection. This global token is then used as a query to cross-attend over patch tokens from all views.
   • Information Restriction: Crucially, OGA employs an information-restricted broadcast mechanism. Instead of adding the cross-attention output as a residual (which would blend normal templates into anomalous regions), it replaces the original patch tokens entirely.
   • Gating: It utilizes temperature-scaled sigmoid gating rather than softmax. This allows for sparse, independent activation of patches, preventing the dilution of anomaly signals by the majority of normal patches and sharpening the distinction between normal and abnormal regions.

2. Multi-view Merging Attention (MMA) for Local Fusion:

   • Goal: To enable fine-grained, token-level communication between views with linear computational complexity.
   • Mechanism: For each view, MMA queries patch tokens from other views using a linear attention kernel (ReLU$^2$). This avoids the quadratic complexity ($O(N^2)$) of standard cross-attention.
   • Adaptive Weighting: The module incorporates a learnable view-importance matrix to weight the contribution of different viewpoints and a token-wise gating mechanism to control the fusion ratio for each token. This ensures that only informative cross-view evidence is integrated, while uninformative or noisy tokens retain their original single-view representation.

Key Contributions

• Identification of Cross-View Information Leakage: The paper defines and analyzes a specific failure mode in multi-view reconstruction-based AD where unconstrained fusion collapses the reconstruction gap.
• GLAD Framework: The first framework to combine vision foundation models with both global (object-level) and local (token-level) cross-view fusion specifically designed for multi-view AD.
• Information-Restricted Design: The introduction of OGA, which replaces rather than adds features, and uses sigmoid gating to prevent the decoder from accessing view-specific normal details that could facilitate "leaked" reconstructions.
• Efficient Linear Fusion: The design of MMA, which achieves scalable local fusion at $O(N)$ complexity through linear attention and adaptive gating, making it suitable for high-resolution features.

Experimental Results
GLAD was evaluated on two benchmarks: Real-IAD (30 object classes, 5 views) and MANTA-Tiny (38 classes, 5 views).

• Performance: GLAD outperforms state-of-the-art single-view and multi-view methods (including Dinomaly and MVAD) across sample-, image-, and pixel-level metrics.
  • On Real-IAD, GLAD achieved the best results on 9 out of 10 metrics, with significant gains in pixel-level precision (P-AP +1.6 over Dinomaly).
  • On MANTA-Tiny, GLAD led on all 10 metrics, showing particular strength in tiny-object localization where single-view features are ambiguous.
• Ablation Studies:
  • Removing MMA primarily hurt sample-level detection, while removing OGA hurt pixel-level localization.
  • The "Information Restriction" components in OGA (token replacement, sigmoid gating, no global feedback to CLS tokens, sharpened temperature) were shown to be critical; removing them caused performance drops, confirming that restricting information flow is necessary to preserve the reconstruction gap.
• Efficiency: The linear attention mechanism in MMA is significantly faster and more memory-efficient than deformable attention or the MVAS block used in prior work, with comparable accuracy.
• Robustness: The model remains stable even when up to three views are masked out, maintaining high sample-level detection scores.

Significance and Claims
The paper claims that effective multi-view fusion in reconstruction-based anomaly detection requires explicit information restriction rather than the unconstrained sharing typical in general multi-view learning. By preventing normal cues from "repairing" anomalous regions, GLAD preserves the reconstruction gap necessary for detection. The authors posit that this principle of information restriction is a fundamental design insight for multi-view reasoning in reconstruction-based settings, offering a new perspective for future work in industrial quality control and anomaly detection.