Mining Deep Part Features for Pedestrian Search

In order to solve the problem of pedestrian's side or occlusions in pedestrian search task, which leads to that neural network can't learn some local features and slow search speed. BatchDrop module for pedestrian re-identification is proposed, which is applied to re-identification stage of two-stage pedestrian search. In the detection stage, faster R-CNN uses ASPP to detect targets. The pedestrian re-identification network is composed of two branches. The global branch extracts the significant global characteristics of pedestrians. Local branch uses BatchDrop module, which randomly discards the same semantically similar part of the feature to enhance the learning of the rest of the features in a training batch. Finally, the feature vectors of the two branches are fused as the final feature vector. The experimental results show that the algorithm has high search accuracy and the search speed is 6-25 times faster than other similar accuracy algorithms.