Extensive comparison of visual features for person re-identification

Person re-identification is one of the most critical tasks in the field of computer vision and has widely applications for abnormal detection and object retrieval in video surveillance. In this paper, we give an extensive comparison for different kinds of visual features including hand-craft features and Convolutional Neural Networks (CNN) features. We run the experiments on three public dataset CASIA, Market1501 and CUHK03. Through A detail comparison and analysis on different features with different similarity measures, we find Colorhistogram and ScalableColor features are most robust to occlusion on CASIA, while GoogleNet and VGGNet features have good robustness as well. For all single features, GoogleNet feature achieves the highest results on Market1501 and CUHK03. For feature fustion, GoogleNet feature with ColorStructure achieve the best result on Market1501 and GoogleNet feature wth Colorhistogram achieve the best result on CUHK03. For similarity measure, Cosine distance is evaluated to be the best one in our experiments.