Reproduce “AA-CLIP: Enhancing Zero-shot Anomaly Detection via Anomaly-Aware CLIP” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Wenxin Ma, Xu Zhang, Qingsong Yao, Fenghe Tang, Chenxu Wu, Yingtai Li, Rui Yan, Zihang Jiang and 1 more
Researchers from USTC and Stanford introduce AA-CLIP, a framework that enhances CLIP's zero-shot anomaly detection capabilities through a two-stage adaptation process, achieving state-of-the-art performance across 11 industrial and medical benchmarks while requiring minimal labeled training data through targeted residual adapters and disentangled text anchors.
Key insights
- Modifying shallow layers of CLIP's text and vision encoders is sufficient for enhancing anomaly awareness
- Separating normal and anomaly embeddings in text space improves discrimination ability
- Lightweight adaptation can maintain CLIP's generalization capabilities while adding task-specific features
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.