Exploring the Impact of Various Contrastive Learning Loss Functions on Unsupervised Domain Adaptation in Person Re-identification

Person Re-identification has drawn great attention in the industrial surveillance system. This paper focuses on the unsupervised domain adaptive case using different contrastive learning loss functions at the source domain and target domain training. Although there are substantial disparities in the data distributions between the source dataset and the target dataset, distinct training strategies on the source dataset continue to exert a significant influence on the ultimate efficacy of Unsupervised Domain Adaptation results, which could be observed in the data distribution of the target dataset before undergoing unsupervised training. This paper systematically conducts the visualization and analysis of the distinct effectiveness of different loss functions on the pre-trained backbone models, especially the clustering quality. Extensive experiments on three large-scale person re-identification datasets achieve commendable results and substantiates the assertions posited in this paper.