Reinforcing Soft Part-Aligned Features for Person Re-identification

Abstract

Aiming to enhance features discrimination for pose variation in person re-identification, reinforcing soft part-aligned features is proposed in this paper. By adding a channel-wise module in two-stream re-identification network, the network can be more robust to person's pose changes. Firstly, the pre-trained GoogLeNet and OpenPose is loaded to extract pedestrian appearance features and pose feature of pre-processed images. Secondly, pose features are fed into the designed channel-wise module. Finally, the two branches are trained jointly by applying bilinear pooling. From the experimental results on the Market-1501 dataset and the DukeMTMC-reID dataset, it shows that reinforcing soft part-aligned features are more discriminative and achieves higher identification accuracy over the state-of-the-arts.