Reproduce “Focus Your Distribution: Coarse-to-Fine Non-Contrastive Learning for Anomaly Detection and Localization” on OpenResearch

About this paper

Ye Zheng, Xiang Wang, Rui Deng, Tianpeng Bao, Rui Zhao, Liwei Wu

SenseTime Research presents a coarse-to-fine non-contrastive learning framework for unsupervised anomaly detection and localization, achieving state-of-the-art performance on industrial datasets like MVTec AD. The method learns exceptionally compact normal feature distributions through a two-stage alignment process and dense pixel-wise non-contrastive learning, effectively identifying subtle, fine-grained defects without synthetic anomaly generation.

Key insights

• The coarse-to-fine alignment strategy effectively handles positional variances and enhances fine-grained anomaly detection by standardizing global context before detailed comparison.
• Dense pixel-wise non-contrastive learning is highly effective in learning compact, discriminative representations of normal data, making subtle deviations from normality detectable.
• The stop-gradient mechanism successfully enables self-supervised learning from purely positive data, acting as a mean-shift algorithm without requiring explicit negative samples.

Summary by alphaXiv.