A Transformer-based Unsupervised Clustering Method for Vehicle Re-identification

Abstract

Current unsupervised re-identification methods use a clustering-based neural network for training. In the vehicle re-identification field, the feature information between different vehicles is small, and it is not easy to distinguish the detailed features of different vehicles using only the basic clustering algorithm for unsupervised learning. When clustering is performed, the general clustering methods inevitably put different vehicles together due to the high similarity. We propose a new re-identification method to solve these problems. This method is based on clustering and use the unsupervised learning. First, we employ the vision transformer structure as a feature extractor. The vision transformer structure can obtain more discriminative and correlated features than the general convolution. Second, we use a fine-grained clustering method to subdivide the clustered information into different vehicles. We trained our method on two open-source datasets, and finally obtained better test results without additional labeling information.