Exploring Sampler Strategies in Unsupervised Person Re-identification Training: Insights and Performance Analysis

Person re-identification holds significant research value within supervised systems characterized by non-overlapping multiple cameras. In recent years, unsupervised learning has made notable strides and has gradually approached the training efficacy of supervised learning. This paper focuses on exploring the influence and analysis of various sampling strategies on overall unsupervised training. We initially delineate a proxylevel memory bank scheme based on camera labels and employ a hard sample mining strategy for selecting negative pairs in a contrastive learning loss. Various sampling strategies, Random sampling, triplet sampling with dissimilar labels, and group sampling yield markedly distinct outcomes across three large-scale datasets, i.e. Market-1501, DukeMTMC-reID, and MSMT17. Detailed analysis and discussion of these results are provided in this study.