Res2Net-based Re-identification Method

The task of re-identification is an important subject in the field of computer vision and is widely used in industry.This paper includes two main contents: first is to explore the effect of the Res2 Net structure on the re-identification task(not limited to pedestrian re-identification),provide a simple and feasible method based on Res2 Net,combined with feature extraction,loss function selection analysis,parameter adjustment method and so on.The other part is the self-proposed Part Attention block for pedestrian tasks,and add this block to the pedestrian re-identification part of our experiment to achieve higher recognition accuracy.Basic knowledge of the Res2 Net is presented in Part2 and the structure of the Res2 Net can be found in 2.2.Part3 introduces the loss function of metric learning.Part4 describes the re-identification tricks we used.Part5 is the core of this paper.We introduce our proposed Res2Net-based reidentification algorithm,including model improvement and loss function selection,as well as the newly proposed Part Attention block for pedestrian re-identification in this paper.Part6 is the experimental part.We used three datasets,including two pedestrian re-identification datasets(Market-1501[14],Duke MTMC-re ID[15])and one packet re-identification dataset(MVB[6]).6.4 introduces the data preprocessing,including different data augmentation approaches to two different kinds of re-identification tasks.Finally 6.5 lists our experimental results.The validity of our proposed Res2 Net re-identification method is verified on three datasets.