GDPS: A general distillation architecture for end-to-end person search

Existing knowledge distillation methods for person search tasks handle detection and re-identification (re-id) tasks separately, which may lead to feature conflicts between the two subtasks. On the one hand, by distilling only the detection task, the network will focus more on the common features of pedestrians, which may affect the performance of re-id. On the other hand, by distilling only the re-id task, the network will be more inclined to focus on the personality characteristics of pedestrians, which may harm the detection performance. To solve this problem, we propose a novel distillation method for person search tasks, treating person search as a single task and distilling different tasks in a unified framework, which is called General Distillation for Person Search (GDPS). Specifically, we optimize the general features of detection and re-id by distilling feature-based knowledge, aiming for accurate localization of individuals. In addition, we focus on the re-id task and perform relationship-based and response-based knowledge distillation to obtain more discriminative person features. Finally, we integrate feature-based, relation-based and response-based knowledge into a general framework to achieve simultaneous distillation of two sub-tasks, which can be readily applied to various end-to-end person search methods. Extensive experiments demonstrate the effectiveness of GDPS across different one-step person search methods. Specifically, AlginPS with ResNet-50 achieves 94.1% in mAP with GDPS on the CUHK-SYSU dataset, which surpasses the baseline 93.1% by 1.0%, and is even better than the ResNet-50 DCN-based teacher model with 94.0% mAP.