Cross-Modality Person ReID with Maximum Intra-class Triplet Loss

Abstract

Compared with conventional ReID task, RGB-IR person re-identification is more significant and challenging owing to enormous cross-modality variations between RGB and infrared images. Most existing cross-modality person re-identification approaches in the infrared and visible domains use the joint training of traditional triplet loss and CE loss to reduce the discrepancy between two modalities. However, as the model gradually converges, the number of positive and negative sample pairs that can be optimized by traditional triplet loss also decreases. In this paper, we adopt a dual-stream CNN structure and propose a maximum intra-class distance to triplet loss (MICT loss) for RGB-IR ReID task. The proposed method enjoys several advantages. First, it can learn specific modal information and shared information better through the CNN structure of the dual-stream branch. Second, it can add the optimization of intra-class cross-modality distance to the traditional triplet loss, which can effectively reduce the intra-class distance and increase the training difficulty of the model. Experimental results demonstrate that our proposed algorithm performs favorably against the state-of-the-art methods that rank-1 accuracy is 61.71% and mAP is 58.06% on SYSU-MM01 dataset.