Unsupervised vehicle re-identification based on mixed sample contrastive learning

This paper proposes a mixed sample contrastive learning framework that constructs memory dictionary with discrete samples and reliable clusters for unsupervised vehicle re-identification. Firstly, we introduce a discrete sample separation (DSS) module including a discrete sample criterion and a discrete sample separation operation. Specifically, for a specific feature cluster, the discrete sample criterion drives the DSS to mine the discrete sample. Such that the original cluster can be separated into a more reliable cluster and discrete samples. Furthermore, a mixed sample contrastive learning (MSCL) strategy is designed to construct a mixed sample memory dictionary for training the model with more superior learning target. Moreover, a discrete sample loss (DSL) is proposed to calculate the contrastive loss of the model and dynamically update the memory dictionary during training. Extensive experiments show our method performs favorably against state of the arts. The code will be published on github after the paper is accepted.