Failure Detection in Image Segmentation under Conditions of Semantic and Covariate Shifts

Whether deep neural networks can provide reliable confidence is of great significance, especially in risk-sensitive scenarios. This work explores the impact of covariate and semantic shifts on segmentation tasks, an area which has not been extensively studied. Covariate shift refers to changes in the data distribution without alterations in the label space, while semantic shift involves changes in both data distribution and label space. We find that model-unknown distributional shifts in test data can transform an overconfidence problem into a situation of making random predictions with arbitrary confidence. The paper proposes a novel approach for effective failure detection that combines holistic image-level analysis and detailed pixel-level information. This approach involves the use of a Gray Level Co-occurrence Matrix (GLCM) to analyze the prediction randomness between adjacent pixels and a Magnitude-Direction Confidence Score Function (MD-CSF) for determining pixel acceptance or rejection. Furthermore, we introduce a new benchmark dataset, the Robot Inspection dataset for Semantic and Covariate shift in Segmentation (RISKS, the homophone of RISCS), to fill the need for datasets capable of evaluating the simultaneous impact of semantic and covariate shifts. Experimental results demonstrate that our method successfully detects image-level failures in segmentation, with MD-CSF outperforming other pluggable CSFs.