GAN-based person search via deep complementary classifier with center-constrained Triplet loss

Abstract

This paper addresses the person search task, which is a computer vision technology that finds the location of a pedestrian and retrieves it in a video taken by a single camera or multiple cameras. This task is much more challenging than the conventional settings for person re-identification or pedestrian detection since the search is susceptible to factors such as different resolutions, similar pedestrians, lighting, viewing angles and occlusion. Moreover, the person search task is a typical big data-small sample problem because each pedestrian only has a few images. It is difficult for the model to learn the discriminant features of pedestrians with a small quantity of pedestrian data. This paper proposes a framework for person search that uses the original training set without collecting extra data by implementing a generative adversarial network (GAN) to generate unlabeled samples. We propose a deep complementary classifier for pedestrian detection to leverage complementary object regions for pedestrian/non-pedestrian classification. In the re-identification section, we propose a center-constrained triplet loss that avoids the complicated triplet selection of the triplet loss and simultaneously pushes away all the distances of rather similar negative centers and the positive center. Experiments show that the GAN-generated data can effectively help to improve the discriminating ability of the CNN model. On the two large-scale datasets, CUHK-SYSU and PRW, we achieve a performance improvement over the baseline CNN. We additionally apply the proposed center-constrained triplet loss and complementary classifiers in the training model, and we achieve mAP improvements over the original method of +1.9% on CUHK-SYSU and +2.5% on PRW.