Leveraging both foreground and background features can substantially enhance the performance of existing out-of-distribution detection methods.
