A cascaded multitask network with deformable spatial transform on person search

Abstract

This article introduces a cascaded multitask framework to improve the performance of person search by fully utilizing the combination of pedestrian detection and person re-identification tasks. Inspired by Faster R-CNN, a Pre-extracting Net is used in the front part of the framework to produce the low-level feature maps of a query or gallery. Then, a well-designed Pedestrian Proposal Network called Deformable Pedestrian Space Transformer is introduced with affine transformation combined by parameterized sampler as well as deformable pooling dealing with the challenge of spatial variance of person re-identification. At last, a Feature Sharing Net, which consists of a convolution net and a fully connected layer, is applied to produce output for both detection and re-identification. Moreover, we compare several loss functions including a specially designed Online Instance Matching loss and triplet loss, which supervise the training process. Experiments on three data sets including CUHK-SYSU, PRW and SJTU318 are implemented and the results show that our work outperforms existing frameworks.