Pedestrian Retrieval Using Valuable Absence Augmentation

Abstract

In this paper, we propose a novel data augmentation method named valuable absence augmentation (VAA) in order to alleviate the overfitting and evaluate the influence of the pedestrian valuable parts for the network performance. Specifically, we first train a base convolutional neural network model and obtain the attention map of the pedestrian. Then, we use the attention map to generate new samples. Finally, original samples and new samples are combined to fine-tune the base network model. We conduct experiments on a large-scale pedestrian retrieval database, i.e., Market-1501. Experimental results show that the pedestrian valuable part has a crucial influence for the network performance and that the proposed method achieves better performance than other state-of-the-art methods.