Class-Wise Thresholding for Robust Out-of-Distribution Detection

Out-of-distribution detectors are usually calibrated with a single global threshold, which degrades badly when the in-distribution label distribution shifts. We set a separate threshold per class and show that this recovers robustness to label shift, improving true positive rate by 20% over a global threshold. The method became the technical backbone of a $100k funded Berkeley DeepDrive proposal on design automation of out-of-distribution data detectors.

IEEE/CVF CVPR Workshops — 2nd Workshop on Fair, Data-Efficient and Trusted Computer Vision (FaDE-TCV)