Ped-Mix: Mix Pedestrians for Occluded Person Re-identification

Shang Gao, Chenyang Yu, Pingping Zhang, Huchuan Lu

Abstract: Occluded person re-identification is a very challenging task due to the interference of occluding objects. Most existing approaches concentrate on modifying the network architecture to facilitate the extraction of more distinctive local features or render the network less sensitive to occlusions. However, it is easy to fail when encountering previously unseen occlusions or when other humans act as occluders, due to the limited occlusion variance in the training set. In this paper, we propose a data augmentation method that blends the target pedestrian with other pedestrians to simulate non-target pedestrian occlusion. Furthermore, we propose a non-target suppression (NTS) loss to reduce the information flow from the occluded region to the final embedding, where the occluded region can be easily obtained from the augmentation. Experimental results demonstrate that this simple augmentation technique yields significant performance improvements in the task of occluded person re-identification.