A Simple Deep Learning Architecture for City-scale Vehicle Re-identification

The task of vehicle re-identification aims to identify a vehicle across different cameras with non overlapping fields of view and it is a challenging research problem due to viewpoint orientation, scene occlusions and intrinsic inter-class similarity of the data. In this paper, we propose a simplistic approach for one-shot vehicle re-identification based on a siamese/triple convolutional architecture for feature representation. Our method involves learning a feature space in which the vehicles of the same identities are projected closer to one another compared to those with different identities. Moreover, we provide an extensive evaluation of loss functions, including a novel combination of triplet loss with classification loss, and other network parameters applied to our vehicle re-identification system. Compared to most existing state-of-the-art approaches, our approach is simpler and more straightforward for training, utilizing only identity-level annotations. The proposed method is evaluated on the large-scale CityFlow-ReID dataset.