Efficient Person Search Via Learning-to-normalize Deep Representation

Ning Lv,Xuezhi Xiang,Xinyao Wang,Jie Yang,Rokia Abdein

DOI: https://doi.org/10.1016/j.neucom.2022.01.028

IF: 6

2022-01-01

Neurocomputing

Abstract:The task of person search is to search for the target pedestrian in the scene images with the given query image, which means to perform detection and re-identification at the same time. In this paper, we propose a one-step method to jointly handle the two subtasks of person search using an end-to-end framework. The proposed framework contains the feature learning module and the multi-task learning module. We explore the performance of learning-to-normalize deep representation on the feature learning module, which allows the network to achieve normalizer selection on each normalization layer. To solve the imbalance problem of hard-to-classify examples, we introduce the focal loss in the multi-task learning module. Besides, we adopt the gradient generalization centralization technique, which makes the training process more stable and enhances the generalization ability of the model. We present extensive experiments on two benchmark datasets, i.e., CUHK-SYSU and PRW. Experimental results prove that our method outperforms the state-of-the-art one-step methods and has a competitive performance compared with two-step methods.(c) 2022 Elsevier B.V. All rights reserved.