Dynamic dual mining framework for long-tailed out-of-distribution detection

Bin Sheng, Dengye Pan, Xiaoqiang Li

Abstract: Detecting out-of-distribution inputs is critical for the reliable and safe deployment of deep learning models in open-world environments. However, most out-of-distribution detection methods rely on the strict assumption of data balance, overlooking the reality that data often follows a long-tailed distribution in real scenarios, which negatively impacts model performance. To overcome this issue, we propose the Dynamic Dual Mining (DDM) framework, which optimally utilizes existing data by performing dual mining on in-distribution data and auxiliary outliers. DDM applies a stronger penalty to hard in-distribution samples and employs prototype-based mining strategy for outliers. Extensive experiments demonstrate that DDM effectively addresses the challenges of long-tailed out-of-distribution detection, achieving state-of-the-art results on CIFAR-10-LT and CIFAR-100-LT, while also exhibiting superior performance on the large dataset ImageNet-200-LT.

External IDs:dblp:journals/apin/ShengPL25