Improving the realism of synthetic images through a combination of adversarial and perceptual losses
Charith Atapattu, Banafsheh Rekabdar · 2019
In recent years, deep learning methods are becoming more widely used; however, large quantities of labeled training data are required for most models. Labeling large datasets is tedious, expensive, and time consuming. Generating large labeled synthetic datasets, on the other hand, is easier and less expensive since annotations are available. But there is usually a large gap between the distribution of the synthetic and real data. In this paper, we propose a novel method based on Generative Adversarial Networks (GANs) to improve the realism of the synthetic images while preserving the annotation information. In our work the inputs of the GANs are synthetic images instead of random vectors. Furthermore, we describe how a perceptual loss can be utilized while introducing the basic features and techniques from adversarial networks for obtaining better results. We evaluate our approach for appearance-based gaze direction classification on the MPIIGaze dataset. The results show that our generated refined images are more realistic and better preserve the annotation information than the refined images generated by the state-of-the-art methods.