Improving person re-identification by background subtraction using two-stream convolutional networks

Abstract

The field of person re-identification is facing problems related to the variation of illumination and background scenes. In order to reduce the impact of those variations, we propose in this work a two-stream re-identification system based on a siamese network (S-CNN). The proposed system takes as input a pair of person images: the original image and the image without background. In the background subtraction step, a segmentation network (SEG-CNN) is used to detect the person body part and capture a complementary information. We experimentally prove that the combination of the two streams (images with and without background) improves the recognition rates. In the rank-1, the improvement is respectively of $$2\%$$ and $$4\%$$ for Market-1501 and DukeMTMC-reID datasets.