Unsupervised Multimodal Anomaly Detection via Stable Reconstruction

Anomaly detection is essential for maintaining product quality and operational safety in industrial environments. However, conventional RGB-based unsupervised methods struggle to capture depth variations and geometric defects, which limits their ability to detect structural anomalies. To address this, we propose a multimodal unsupervised anomaly detection framework that integrates RGB images and 3D information through a stable reconstruction module designed to suppress unstable reconstruction errors in normal regions and enhance the reconstruction gap between normal and anomalous samples. Extensive experiments on the MVTec 3D-AD and Eyecandies datasets demonstrate that the proposed multimodal fusion approach consistently outperforms single-modality baselines in terms of AUROC and AUPRO.