Multi-dimensional exemplar-based texture synthesis.

摘要

Given an image or video of a scene, this dissertation describes methods to generate larger images and longer videos, which are similar to the respective inputs. This problem is known in computer graphics as texture synthesis, and has previously been studied mainly for image textures. In this dissertation I describe several new techniques that allow synthesis of not only textures, but also images with stronger content structure, general video and character animations.; This thesis explores three different representations. Video textures are infinitely playing full-frame videos which are generated by rearranging frames copied from the source video. Video sprites are a special form of video textures, for which a bitmap with opacity and velocity information (a “sprite”) is extracted from video and used to generate an animation. Both approaches treat the problem of assembling the output from the input only in the time dimension; the image or sprite itself is always copied as a whole from the input. The third representation, multi-dimensional textures, which includes both image and video textures, is generated by assembling arbitrary regions in multiple dimensions using graph cuts. For video sequences, this allows assembling each output image from parts taken from multiple input images.