Realistic view synthesis of a structured traffic environment via adversarial training

Generating a realistic image from a novel viewpoint has always been a key problem in image-based rendering and other related domains. In this paper we utilize the state-of-the-art generative adversarial networks(GAN) to synthesize novel views of a structured scene. Based on our proposed representations for traffic scene, a realistic image of a certain viewpoint can be generated via conditional GANs, given the geometric layout of the corresponding position and pose. In order to preserve the geometric property of the input in the generated image, we propose a simple but effective constraint in the generator network. Qualitative and comparative results have validated our method as well as shown its effectiveness and efficiency.