Fine-grained alignment network and local attention network for person re-identification

Due to the influence of person posture changes, light angle of view, background and other factors, person re-identification is a challenging task. To improve the identification accuracy, recent studies have divided the pedestrians in the dataset into several blocks to extract the local features of the image for re-identification. However, these methods have such problems as the mismatch of local features of the human body and the loss of contextual clues of non-human body parts. To solve the above problems, this paper proposes a partially aligned network that can be used for person re-identification, which uses accurate local features to increase the ability of human body semantic parsing to model arbitrary contours. On this basis, the local attention network captures contextual cues that are not part of the human body. In addition, by aligning the local features of human body semantic parsing, the robustness and mobility of the model can be effectively increased. The experimental results obtained with the three datasets, Market-1501, DukeMTMC and CUHK03, show the effectiveness of the proposed model.