Abstract
The goal of this work is to synthesize high-resolution panoramic images of a scene. We propose a system that takes low-resolution panoramic video frames as the input, extracts geometric information as the reference, and produces high-resolution panoramas of the scene as the output. The deep learning module of our system benefits from the structure-from motion geometry and learns to decouple the content and the style of an input low-resolution patch for hallucinating its high-resolution version. We show that the high-resolution panoramas generated by our system achieve a better quality of detail and color enhancement than those produced by existing super-resolution and style-transfer methods.