Diverse Knowledge Distillation for End-to-End Person Search

• DOI: 10.1609/aaai.v35i4.16454
• Publication Date: 2022-09-08T18:19:51Z
• Authors: Xinyu Zhang; Xinlong Wang; Jia-Wang Bian; Chunhua Shen; Mingyu You
• Keywords: End-to-end principle

Person search aims to localize and identify a specific person from gallery of images. Recent methods can be categorized into two groups, i.e., two-step end-to-end approaches. The former views as independent tasks achieves dominant results using separately trained detection re-identification (Re-ID) models. latter performs in an fashion. Although the approaches yield higher inference efficiency, they largely lag behind those counterparts terms accuracy. In this paper, we argue that gap between kinds is mainly caused by Re-ID sub-networks methods. To end, propose simple yet strong network with diverse knowledge distillation break bottleneck. We also design spatial-invariant augmentation assist model invariant inaccurate results. Experimental on CUHK-SYSU PRW datasets demonstrate superiority our method against existing -- it par accuracy state-of-the-art while maintaining high efficiency due single joint model. Code available at: https://git.io/DKD-PersonSearch.