Mixture Outlier Exposure for Out-of-Distribution Detection in Fine-grained Settings

Enabling out-of-distribution (OOD) detection for DNNs is critical for their safe and reliable operation in the open world. Despite recent progress, current works often consider a coarse level of granularity in the OOD problem, which fail to approximate many real-world fine-grained tasks where high granularity may be expected between the in-distribution (ID) data and the OOD data (e.g., identifying novel bird species for a bird classification system in the wild). In this work, we start by carefully constructing four large-scale fine-grained test environments in which existing methods are shown to have difficulties. We find that current methods, including ones that include a large/diverse set of outliers during DNN training, have poor coverage over the broad region where fine-grained OOD samples locate. We then propose Mixture Outlier Exposure (MixOE), which effectively expands the covered OOD region by mixing ID data and training outliers, and regularizes the model behaviour by linearly decaying the prediction confidence as the input transitions from ID to OOD. Extensive experiments and analyses demonstrate the effectiveness of MixOE for improving OOD detection in fine-grained settings.