Pedestrian Retrieval via Part-Based Gradation Regularization in Sensor Networks

In this paper, we propose a novel label distribution approach named part-based gradation regularization (PGR) for pedestrian retrieval in sensor networks. Considering different importance of various body parts, we present a gradual function to assign pedestrian label for each horizontal part. In this way, we can conduct part-based supervised learning using the identification network. The proposed PGR not only learns the discriminative local convolutional neural network-based features, but also considers the significance of assigning pedestrian label for different horizontal parts. Experimental results show that the proposed PGR obtains better performance than other approaches on three pedestrian retrieval databases, i.e., Market-1501, CUHK03, and DukeMTMC-reID databases.