Fusional Modality and Distribution Alignment Learning for Visible-Infrared Person Re-Identification

Abstract

The Visible-Infrared Person Re-Identification (VI-ReID) task aims to retrieval pedestrian images with the same labels across different modalities. VI-ReID is a very challenging task due to the huge intra-modality variation and cross-modality gap. Existing methods are mainly based on the feature alignment to mitigate the modality's gap, however, using only feature-level constraints does not mitigate cross-modality gap well. We propose a fusional modality and distribution alignment learning network (FMADALNet) to mitigate modality's gap and align modality's distribution to learn modality-shared feature representations. FMADALNet contains a lightweight fusional modality generation module (FMGM). FMGM constructs a fusional modality that incorporates heterogeneous image features and contains only modality-shared information to mitigate modality gap at the pixel-level. In addition, to mitigate the differences in the distribution of the different modalities, we design a Hetero-center Maximum Mean Discrepancy loss (HcMMD), which reduces the differences in the distribution of the different modalities in a displaying manner. Extensive experimental results on two public datasets show that our proposed method achieves impressive performance compared to state-of-the-art methods.