An Empirical Study on Face Re-identification Using Machine Learning Models

Person Re-Identification (ReID) aims to match the same person across different camera views, and it is a core task in intelligent security and video analysis. This study is based on comparative experiments of various model architectures, loss functions, and hyperparameter configurations, and comprehensively analyzes their impacts on performance. The experiments cover models such as Cross-Entropy (CE) and its variants, DenseNet-121, ResNet50, and SwinV2, as well as the optimization effects of key parameters like Batch Size, the data augmentation strategy (erasing p) and the warm-up period (warm epoch). All experimental data are strictly cited, and the conclusions are derived based on the actual results without any subjective modification.