iCAS: A In-Context Anomaly Segmentation Framework for Industrial Visual Inspection

ICLR 2026 Conference Submission16223 Authors

Keywords: Anomaly Detection, In-Context Learning

Abstract: Visual in-context prompting has recently made promising progress, achieving training-free segmentation with a generalized model derived from large-scale pre-training. However, we observe that these in-context segmentation models fail on the anomaly detection task, e.g., visual inspection. In this study, we propose iCAS, a novel model for In-Context Anomaly Segmentation enabling automatic defect annotation and visual prompting anomaly segmentation. The framework is built upon an in-context mask transformer, further enhanced by a greedy query selection strategy and a mask-level feature matching module to improve both sensitivity and generalization. Further, we propose the General-to-Specific pre-training to solve the weak generalization problem caused by the scarcity of anomalous samples. Finally, we conduct comprehensive experiments under a variety of anomaly detection and segmentation tasks. Evaluations on multiple publicly available datasets show the generalization and effectiveness of our method.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 16223