Reproduce “InCTRLv2: Generalist Residual Models for Few-Shot Anomaly Detection and Segmentation” on OpenResearch

About this paper

Jiawen Zhu, Mengjia Niu, Guansong Pang

A generalist residual model, InCTRLv2, advances few-shot anomaly detection and segmentation by combining in-context residual learning with semantic-guided dual-branch modules leveraging Vision-Language Model priors. This framework demonstrates superior generalization and more precise anomaly localization across diverse unseen domains, achieving improvements of up to 3.9% AUROC for image-level detection on industrial datasets and 16.3% AUROC for pixel-level detection on medical datasets compared to prior methods.

Key insights

• Integrating VLM text priors with in-context residual learning provides a deeper, semantic-guided understanding of anomalies, addressing limitations of purely visual discrepancy modeling.
• A dual-branch architecture, with DASL focusing on discriminative abnormality and normality, and OASL on domain-stable normality, offers complementary and robust anomaly perception.
• Explicitly modeling intrinsic normality patterns through the OASL branch improves generalization by offering a stable reference for anomaly correction and reducing false positives.

Summary by alphaXiv.