Understanding and Overcoming Cross-modal Fusion Bias in Multimodal Anomaly Detection From A Fisher Information Perspective

2026-08-02

AI Summary

This work addresses the performance bottleneck in multimodal anomaly detection caused by cross-modal fusion bias. It introduces, for the first time, the Fisher Information Matrix to quantify this bias and proposes an Unbiased Cross-Modal Fusion Framework (UCFB) that enables plug-and-play bias calibration. UCFB leverages Fisher information-guided dynamic regularization and Canonical Correlation Analysis to adaptively adjust fusion weights, thereby mitigating inter-modal inconsistencies. Evaluated on RGB-D multimodal data, the method consistently achieves significant performance gains across one-class, multi-class, and few-shot settings on the MVTec 3D-AD and Eyecandies benchmarks.

Abstract

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.

Problem

• Multimodal Anomaly Detection
• Cross-modal Fusion Bias
• Fisher Information
• RGB-D Fusion
• Performance Bottleneck

Innovation

• Fisher Information
• Cross-modal Fusion Bias
• Multimodal Anomaly Detection
• Dynamic Calibration
• Canonical Similarity Analysis