Single Image Based Metric Learning via Overlapping Blocks Model for Person Re-Identification

Yipeng Chen, Cairong Zhao, Tianli Sun; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2019, pp. 0-0

Considering the pedestrian structure characteristics, the first step of many person re-identification algorithms is to divide the pedestrian images or feature map into several blocks, and then the blocks in the same location are used to calculate the special loss functions that metric the differences between different images, to reduce the distance between intra-samples and to increase the distance between inter-samples. However, most of those blocks based deep metric learning methods only measure the difference between different images, but ignored the metrics between different blocks in a single image. In this paper, we propose a novel blocks based method for person re-identification called Overlapping Blocks Model (OBM), in which an innovative strategy of overlapping partition on convolutional features is used to construct multiple overlapping blocks structure and a novel overlapping blocks loss function is utilized to measure the difference between different blocks in a single image, to ensure more blocks can bring more discriminate information and higher performance. We conduct thorough validation experiments on the Market-1501, CUHK03, and DukeMTMC-reID datasets, which demonstrate that our proposed Overlapping Blocks Model can effectively improve the recognition performance of networks by adding the multiple overlapping blocks structure and the overlapping blocks loss.