Two-stage metric learning for cross-modality person re-identification

Cross-modality person re-identification faces big challenges as the different characteristics of images collected by visible and thermal cameras. The existing deep learning methods always use metric learning to learn the discriminative features. However, the existing metric learning is executed based on batch examples, the solution is local optimal. In order to learn a global solution, we propose a two-stage metric learning (TML) method, which uses local and global metric learning successively. In the first stage, a local metric learning is used based on mini-batch images via triplet loss. A new mixed-modality triplet loss is proposed to train more valid triplet examples. It supervises to learn more efficient features for the next stage. In the second stage, a global metric learning is adopted based on the features of all training images. Experiments are conducted on the public SYSU-MM01 dataset. The TML achieved 39.75% in Rank-1 and 42.73% in mAP, which surpass the state-of-the-art performance.