Augmented Reality Image Generation With Optical Consistency Using Generative Adversarial Networks

Various methods are used for estimating light source informations from real objects to achieve optical consistency in augmented reality (AR), but in practice, there are difficulties in using real objects. We propose a method of achieving optical consistency without estimating the light source information, using generative adversarial networks (GANs) that input AR images without optical consistency and mask image. The generated AR images from our proposed method show the appropriate expression of drop shadows and reflections of surrounding objects.