Multi-domain Cross-dataset and Camera Style Transfer for Person Re-Identification

The major reason of performance drop in person reidentification (re-id) models is domain gap between various datasets. Each person re-id dataset is different from other dataset because of change in lightning conditions, camera angle, pose, illumination, background and resolution. This results in domain gap between these datasets and leads to performance drop in person re-identification models. The domain gap not only exits between distinct datasets but it is also present in a dataset which contains images taken from multiple cameras. Generative Adversarial Networks (GANs) have achieved successful results in various fields, especially for image-to-image translation. In this paper, we propose an unsupervised domain adaptation method based on ComboGAN and Xception to perform cross-dataset and camera style translation. Our framework generates images in different styles of a single person according to various domains. Experimental results show the effectiveness of the proposed model on two popular re-id datasets Market-1501 and PETA.