Subtask-dominated Supervised Pretraining Transfer Learning for Person Search

Chuang Liu (Shanghai Jiao Tong University),* Hua Yang (Shanghai Jiao Tong University), Shibao Zheng (SJTU)
The 33rd British Machine Vision Conference


Since person search datasets are of limited scale due to expensive efforts to collect large-scale annotated datasets, existing one-step person methods leverage models pretrained on ImageNet to overcome this shortcoming. However, pretraining on ImageNet suffers from the large domain gap between ImageNet and target datasets. To address this issue, we propose a Subtask-dominated Supervised Pretraining (SSP) transfer learning method. The proposed SSP method takes the person re-identification (Re-ID) subtask as the dominant subtask of one-step person search and pretrains the backbone model in the Re-ID subtask with annotated data. The pretrained backbone weights can provide the one-step person search model with a better initialization to help it converge to a better solution. Specifically, the proposed SSP method surpasses the ImageNet pretraining method by 6.6\% mAP and 1.9\% top-1 on the PRW dataset. Besides, to reduce the impact of person detection subtask on the dominant Re-ID subtask, we further design a Multi-level RoI Fusion Pooling layer to enhance the discrimination ability of learned person features for one-step person search. Extensive experiments on the PRW and CUHK-SYSU datasets demonstrate the superiority and effectiveness of the proposed method.



author    = {Chuang Liu and Hua Yang and Shibao Zheng},
title     = {Subtask-dominated Supervised Pretraining Transfer Learning for Person Search},
booktitle = {33rd British Machine Vision Conference 2022, {BMVC} 2022, London, UK, November 21-24, 2022},
publisher = {{BMVA} Press},
year      = {2022},
url       = {}

Copyright © 2022 The British Machine Vision Association and Society for Pattern Recognition
The British Machine Vision Conference is organised by The British Machine Vision Association and Society for Pattern Recognition. The Association is a Company limited by guarantee, No.2543446, and a non-profit-making body, registered in England and Wales as Charity No.1002307 (Registered Office: Dept. of Computer Science, Durham University, South Road, Durham, DH1 3LE, UK).

Imprint | Data Protection