Decomposing Texture and Semantics for Out-of-distribution Detection

Out-of-distribution (OOD) detection has made significant progress in recent years because the distribution mismatch between the training and testing can severely deteriorate the reliability of a machine learning system.Nevertheless, the lack of precise interpretation of the in-distribution limits the application of OOD detection methods to real-world system pipielines. To tackle this issue, we decompose the definition of the in-distribution into texture and semantics, motivated by real-world scenarios. In addition, we design new benchmarks to measure the robustness that OOD detection methods should have. To achieve a good balance between the OOD detection performance and robustness, our method takes a divide-and-conquer approach. That is, the model first tackles each component of the texture and semantics separately, and then combines them later. Such design philosophy is empirically proven by a series of benchmarks including not only ours but also the conventional counterpart.