Progressive Corruption-Invariant Feature Mining for Robust Visible-Infrared Person Re-Identification

Device abnormality, signal transmission damage, inferior weather and image compression loss cause significant performance degradation of existing Visible-infrared person Reidentification (VI-ReID) methods. Most of the previous works focus on narrowing the large modality discrepancy, but little attention is paid to the models robustness to various corrupted images, which is a vital aspect in real-world scenes. We propose a unique Progressive Corruption-Invariant Feature Mining (PCFM) strategy to solve this problem. To enforce the model to adapt to varied corrupted images during training, we suggest a Random Patch Corruption Augmentation (RPCA), which randomly adds various levels of corruption patches. Then we introduce a Progressive Global-Local Feature Learning strategy which consists of a shared transformer with graph and a Bimodal Three-Stage Learning (BTL). This strategy is employed to learn the robust, modality-shared, global-local pedestrian features gradually in stages. Moreover, a Hard Boundary Loss (HBL) is designed to achieve batch-level global aggregation and align the corrupted features of distinct modalities. Extensive experiments are conducted on SYSU-MM01, RegDB, SYSU-MM01-C, RegDB-C datasets. The experiment results prove that our proposed method can effectively promote the recognition performance for corrupted images, which becomes more robust in real-world applications.