Cross-Modality Person Retrieval with Cross-Modality Loss Functions

Cross-modality person retrieval aims to search a pedestrian image of RGB modality from infrared (IR) pedestrian images and vice versa. Many approaches utilize dual-stream models to extract the features of images from different modalities. Then, they use different types of loss functions to overcome the intra-class modality variations and the large cross-modality discrepancy. In this review, we introduce three types of loss functions used in cross-modality person retrieval including probability based, distance-based, and center-based loss functions. Afterwards, we combine the baseline network with different loss functions, and perform the experiments on a publicly available dataset, i.e., SYSU-MM01 to evaluate their impact on cross-modality person retrieval.