Multi-stage attention and center triplet loss for person re-identication

摘要

Person re-identification(Re-ID) has been a hot topic in the field of computer vision for the past few years. In order to solve the problem of misalignment between different images, most of the existing algorithms use a certain method to manually divide the image into several parts(such as uniform block, segmentation, and pose estimation model), and then extract local features through a multi-branch network structure. However, the compulsory region division does not give full play to the advantages of the deep learning network’s automatic feature extraction, and will bring additional calculations. This paper designs the architecture of Multi-stage Attention which achieves the goal of automatically extracting the discriminative features with a single branch structure. Meanwhile, there are few such studies that enhance the power of discriminating features by designing loss functions. Triplet loss, one of the most commonly used loss functions, suffers from difficulty of mining hard triplets and time consuming. To address this issue, we propose an innovative loss function, namely Center Triplet Loss(CTL), to learn a center of each class and to find the closest negative sample to the center. The triplet of CTL consists of an anchor and the corresponding center, and the closest negative sample. As a result, the model is easy to train and stable because the sample pairs are no longer randomly selected. Our algorithm(MACTL: Multi-stage Attention and Center Triplet Loss for Person Re-Identication) outperforms state-of-the-art(sota) on the datasets of Market-1501 and DukeMTMC-reID, with less branches and more stable.