Exploiting Improved Cyclic Generative Adversarial Networks to Generate Images from Different Angles of Views

Abstract

Virtual Reality (VR), Augmented Reality (AR), and other similar technologies are becoming increasingly popular nowadays. However, despite the increasing number of related devices being launched, the feedback tends to be disappointing. A significant contributor to this dissatisfaction is the blurred viewpoints and unclear levels, which often lead to a suboptimal user experience. To address this issue, this research focuses on training a modified version of the Cycle-Consistent Generative Adversarial Networks (CycleGAN) model to generate an image from another angle of view based on a given image. This work has fine-tuned specific layers of the CycleGAN to better suit the requirements of this work. These improvements give the model enhanced adaptivity in handling visual image transformations. Consequently, the proposed model achieves superior results compared to other models, with minimal image distortion. This is likely due to the minimal angular difference between the generated images. In the future, it is expected to expand the number of viewpoints generated and enhance the model's efficiency in processing images with varying resolutions.