Hierarchical distillation learning for scalable person search. (June 2021)

• Record Type: Journal Article
• Authors: Li, Wei; Gong, Shaogang; Zhu, Xiatian
• Is Part Of: Pattern recognition. Volume 114(2021)
• Publication Date: 2021-06
• Subjects: Person search -- Person re-identification -- Person detection -- Knowledge distillation -- Scalability -- Model inference efficiency; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2021.107862
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: We investigate for the first time the scalability problem involved in person search. This is a fundamentally significant problem to be solved for scaling up the deep learning solutions to person search in the real-world applications. We formulate a hierarchical distillation learning (HDL) approach for more discriminating knowledge transfer from a stronger teacher model into an efficient student model. We design a simple and effective teacher model for joint learning of person search, which largely facilitates the knowledge distillation by avoiding knowledge transfer between structure inconsistent teacher and student models. We demonstrate the model cost-effectiveness and performance advantages of our HDL over the state-of-the-art alternative approaches on three person search benchmarks. Abstract: Existing person search methods typically focus on improving person detection accuracy. This ignores the model inference efficiency, which however is fundamentally significant for real-world applications. In this work, we address this limitation by investigating the scalability problem of person search involving both model accuracy and inference efficiency simultaneously. Specifically, we formulate a Hierarchical Distillation Learning (HDL) approach. With HDL, we aim to comprehensively distil the knowledge of a strong teacher model with strong learning capability to a lightweight student model with weak learning capability. To facilitate the HDL process, we design a simple and powerful teacher model for joint learning of person detection and person re-identification matching in unconstrained scene images. Extensive experiments show the modelling advantages and cost-effectiveness superiority of HDL over the state-of-the-art person search methods on three large person search benchmarks: CUHK-SYSU, PRW, and DukeMTMC-PS.