StRDAN: Synthetic-to-Real Domain Adaptation Network for Vehicle Re-Identification

Sangrok Lee, Eunsoo Park, Hongsuk Yi, Sang Hun Lee

Abstract: Vehicle re-identification (Re-ID) aims to obtain the same vehicles from vehicle images. Vehicle Re-ID is challenging but important for analyzing and predicting traffic flow in the city. Deep learning methods have achieved huge progress in this task. However, requiring a large amount of data is a critical shortcoming. To tackle this problem, we explore the method that uses inexpensive synthetic data to improve performance. Inspired by domain adaptation and semi-supervised method, we propose joint and disjoint losses that fully utilize the synthetic data and their labels without extra cost. We evaluate our network on VeRi and CityFlow dataset with mean average precision (mAP) metric. The results show that our method outperforms the real-world data only baseline by up to 12.87% in CityFlow and 3.1% in VeRi