Reproduce “UniVAD v2: Unified Visual Anomaly Detection via Support-Conditioned Boundary Construction” 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
Zhaopeng Gu, Bingke Zhu, Zhaowen Li, Guibo Zhu, Yingying Chen, Ming Tang, Peng Su, Jinqiao Wang
UniVAD v2 presents a unified visual anomaly detection framework that constructs adaptive decision boundaries by integrating both normal and optional abnormal support sets. The framework achieves a mean image-level AUC of 85.7% and pixel-level AUC of 93.1% across six diverse datasets under a 1N+1A-shot protocol, demonstrating effective inference-time boundary customization for varying anomaly tolerance.
Key insights
- Jointly modeling both local (retrieval) and global (Optimal Transport) support-query relations is crucial for comprehensive normal-side boundary construction, effectively detecting both local defects and structural/logical anomalies.
- Explicitly incorporating abnormal references as "rejection-side evidence" through a residual formulation allows dynamic and controllable adjustment of decision boundaries at inference time without requiring model retraining.
- An adaptive coordination mechanism that weighs complementary normal-side evidence sources based on their reliability for each specific episode significantly enhances overall anomaly detection performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.