Whole constraint and partial triplet-center loss for infrared-visible re-identification

In this paper, we propose a hybrid cross-dual-path feature learning network (HCDFL) to learn local discriminative characteristics and exploit a novel loss function called whole constraint loss and partial triplet-center loss (WCPTL) to narrow the discrepancy between cross-modality. HCDFL firstly extracts person features of two modalities, respectively, then horizontally slices the extracted features into p stripes and maps them into the shared space, finally utilizes modality-specific identity loss, cross-entropy and WCPTL proposed in this paper to improve the overall performance. Within WCPTL, whole constraint loss is utilized to narrow the discrepancy between two modalities, and partial triplet-center loss is utilized to ensure that samples from the same category can be close to the category center and away from other categories center. Experimental results demonstrate that the proposed method achieves better performance than other methods on two public datasets SYSU-MM01 and RegDB.