Stereo Images Generation from a Single Image Based on the View Synthesis Network

• Main Authors: Lo, Yuan-Mau, 羅元懋
• Other Authors: Shih, Zen-Chung
• Format: Others
• Language: en_US
• Published: 2018
• Online Access: http://ndltd.ncl.edu.tw/handle/xdr3jm

碩士 === 國立交通大學 === 多媒體工程研究所 === 107 === Although there are a few methods generating stereo images from a single image, they are based on the traditional warping method, which considers only the translation of pixels in the image. But in the real world, when gazing an object, the two views of human eyes present different faces of it due to the distance between the two eyes. The traditional warping method cannot produce a sufficient stereoscopic sense, and the distortion may even happen in the results. Therefore, in this thesis, we propose a system which can generate stereo images from a single image considering both translation and rotation of objects in the image. Our modified appearance flow network is more general and suitable for our system. We also use the reference image to improve the inpainting method. The quality of our results is better than those using the traditional warping. Our results can better keep the structure of objects in the input image. In addition, our system does not limit the size of the input image. Most importantly, due to considering the rotation of objects, our results are more stereoscopic while watching them with a device.