Regional Feature Learning using Structural Analysis of Attributes in a Bipartite Attention Framework for Vehicle Re-Identification

Abstract

Abstract Vehicle re-identification aspires to discover the target vehicle from the images obtained from the multiple non-overlapping real-world surveillance cameras. The challenges of illuminations, pose variations, resolution and various other attributes adds to the difficulty in re-identification. To overcome these difficulties, fine-grained visual differences in the vehicles are recognized for accurate predictions. Along with the coarse-grained features such as vehicle colour, model the customized features like annual service signs, logo, hangings are additionally considered. In this research work, we have designed a new dataset named Attributes27 with 27 classes of hierarchically labelled attributes to work on our proposed bipartite attention framework. The bipartite framework comprises of three blocks: A double branched convolution neural network layer extracts the global and local features in each vehicle image. Secondly, the self-attention block attached to each branch to detect the Region of Interests (ROIs). And finally the partition-alignment block is deployed to obtain the regional features from the gathered ROIs. Attributes27 and VeRi-776 datasets are the two datasets on which our proposed system is evaluated and the results show significant improvement in accuracy and also concentrate on the specific regional attributes present in each vehicle. The proposed method exhibits a performance of 98.5% and 84.3% respectively on Attributes27 and VeRi-776 datasets which is compared to be higher than the existing methods which exhibited 78.6% of accuracy.