A Softened Similarity with Global Attention Method for Unsupervised Person Re-Identification

There are many factors that affect the accuracy of the algorithm in the field of unsupervised person re-identification. This article focuses on two of them. One is the hard quantization loss caused by the clustering algorithm, and the other is the interference problem caused by the noise information in the image background. In this paper, we propose a method to solve the above problems. Our method consists of two parts. One is to use a non-clustering algorithm that softens similarity to process tags to minimize the impact of hard quantization losses. The other is to add global relational attention to calculate the degree of association between each feature point and the whole world. In this way, the attention value is calculated, and the foreground and background of the image are separated from the space and the channel, so that the algorithm focuses on the foreground information and try to avoid the interference caused by the background noise information. We use market-1501 and Duke data sets to conduct experiments. The experimental results show that our algorithm is significantly better than other advanced algorithms.