Generation of Artificial Training Data for Deep Learning
Can artificial training data be used for deep learning applications in computer vision? We investigated this by building a framework that uses computer graphics to generate large quantities of images portraying humans. Generative Adversarial Networks (GANs) were used to bridge the gap in appearance between the generated and real images. A dataset's potential was estimated by first using it to train a person re-identification model, and then evaluating the model on real images. Results showed that datasets that had been put through a GAN had higher potential than those which had not. We were not able to replace real images with artificial ones, but our results showed promise for further work in substituting and complementing real images.