Vehicle Re-Identification using Shifted Window Transformer

Abstract

Vehicle re-identification has become a key area in computer vision which has im portant applications in areas like vehicle tracking, commodity tracking, security tracking, etc. But Vehicle re-identification inherits two challenges; there are large variations between the similar classes, similar classes have different looks in dif ferent viewpoints, and slight differences between different classes, two different vehicles may look the same as companies are making vehicles with similar ap pearances. Various CNN-based Vehicle Re-ID have been proposed but CNN faces two main challenges; only contain the information of neighborhood and loss of information due to sampling and convolutions. Transformer have recently been introduced in vision community and achieving superior results because they can solve fundamental problems of CNNs by keep long term relation and not losing any information. A novel approach based on Shifted Window Transformer was introduced that in this thesis that uses transformers model to tackle the vehicle re-id problem and uses side information module to enhance the discrimination ability of the model. To the best of our knowledge Shifted Window Transformer have first-time been used as a backbone in vehicle re-identification task and results were promising, the model was evaluated on benchmark dataset VeRi.