Vehicle Re-Identification based on Ensembling Deep Learning Features including a Synthetic Training Dataset, Orientation and Background Features, and Camera Verification

Abstract

Vehicle re-identification has the objective of finding a specific vehicle among different vehicle crops captured by multiple cameras placed at multiple intersections. Among the different difficulties, high intra-class variability and high inter-class similarity can be highlighted. Moreover, the resolution of the images can be different, which also means a challenge in the re-identification task. Intending to face these problems, we use as baseline our previous work based on obtaining different deep learning features and ensembling them to get a single, stable and robust feature vector. It also includes post-processing techniques that explode all the information provided by the CityFlowV2-ReID dataset, including a re-ranking step. Then, in this paper, several newly included improvements are described. Background and orientation similarity matrices are added to the system to reduce bias towards these characteristics. Furthermore, we take into account the camera labels to penalize the gallery images that share camera with the query image. Additionally, to improve the training step, a synthetic dataset is added to the original one.