Part-based pyramid loss for person re-identification

Person re-identification (ReID) is a challenging problem in computer vision, meanwhile attracted the attention of industry. Person ReID focuses on identifying person among multiple different cameras. A key under-addressed problem is to learn a good metric for measuring the similarity among images. Recently, deep learning networks with metric learning loss has become a common framework for person ReID, such as triplet loss and its variants. However, the previous method mainly uses the distance to measure the similarity and the distance measure is more sensitive when the scale changes. In this paper, we propose part-based pyramid loss to learn better similarity metric for the person ReID, in which batches of quadruplet samples as the input. Specifically, we simultaneously use the relationship of distance and angle among samples learn the local body-parts features of person images. Our approach uses the pyramid relationship in triangles as a measure of similarity, minimising the angle at the negative point of the triangle. Pyramid loss can learn better similarity metric and achieve a higher performance on the person ReID benchmark datasets. The experimental results show that, our method yields competitive accuracy with the state-of-the-art methods.