Spatial Invariant Person Search Network

摘要

A cascaded framework is proposed to jointly integrate the associated pedestrian detection and person re-identification in this work. The first part of the framework is a Pre-extracting Net which acts as a feature extractor to produce low-level feature maps. Then a PST (Pedestrian Space Transformer), including a Pedestrian Proposal Net to generate person candidate bounding boxes, is introduced as the second part with affine transformation and down-sampling models to help avoid the spatial variance challenges related to resolutions, viewpoints and occlusions of person re-identification. After further extracting by a convolutional net and a fully connected layer, the resulting features can be used to produce outputs for both detection and re-identification. Meanwhile,we design a directionally constrained loss function to supervise the training process. Experiments on the CUHK-SYSU dataset and the PRW dataset show that our method remarkably enhances the performance of person search.