Decomposing texture and semantic for out-of-distribution detection[Formula presented]

Abstract

The out-of-distribution (OOD) detection task assumes samples that follow the distribution of training data as in-distribution (ID), while samples from other data distributions are considered OOD. In recent years, the OOD detection tasks have made significant progress since many studies observed that the distribution mismatch between training and real datasets can severely deteriorate the reliability of AI systems. Nevertheless, the lack of precise interpretation for the in-distribution (ID) limits the application of the OOD detection methods to real-world systems. To tackle this, we decompose the definition of the ID into texture and semantics, motivated by the demands of real-world scenarios. We also design new benchmarks to measure the robustness that OOD detection methods should have. Our proposed benchmark verifies not only the precision but also the robustness of the detection models. It is crucial to measure both factors in OOD detection as they indicate different traits of the model. For instance, precision is relevant to scenarios that detect minor cracks in the conveyor belt of a smart factory, whereas robustness pertains to maintaining performance under diverse weather conditions, as required by autonomous driving. To achieve a good balance between the OOD detection performance and robustness, our method takes a divide-and-conquer approach. Specifically, the proposed model first handles each component of the texture and semantics separately and then fuses these later. This philosophy is empirically proven by a series of benchmarks including both the proposed and the conventional counterpart. By decomposing the prior “unclear” definition of the ID into texture and semantic components, our novel approach better suits the demands of a reliable machine learning system, which requires robustness and consistent performance across varied scenarios. Unlike prior works, our approach does not rely on any extra datasets or labels. This prevents our proposed framework from being dependent on a particular dataset distribution.