Abstract

Video texture, a new type of medium, can produce a new video with a continuously varying stream of images from a recorded video. It is synthesized by reordering the input video frames in a way which can be played without any visual discontinuity. However, video texture still experiences few unappealing drawbacks. For instance, video texture techniques can only generate new videos by simply rearranging the order of frames in original videos. Therefore, all the individual frames are the same as before and the result would suffer from "dead-ends" if the current frame could not discover similar frames to make a transition. In this thesis, we propose several new approaches for synthesizing video textures. These approaches adopt dimensionality reduction and regression techniques to generate video textures. Not only the frames in the resulted video textures are new, but also the "Dead end" problem is avoided. First, we have extended die work of applying principal components analysis (PCA) and autoregressive (AR) process to generate video textures by replacing PCA with five other dimension reduction techniques. Based on our experiments, using these dimensionality reduction techniques has improved the quality of video textures compared with extraction of frame signatures using PCA. The synthesized video textures may contain similar motions as the input video and will never be repeated exactly. All frames synthesized have never appeared before. We also propose a new approach for generating video textures using probabilistic principal components analysis (PPCA) and Gaussian process dynamical model (GPDM). GPDM is a nonparametric model for learning high-dimensional nonlinear dynamical data sets. We adopt PPCA and GPDM on several movie clips to synthesize video textures which contain frames that never appeared before and with similar motions as original videos. Furthermore, we have proposed two ways of generating real-time video textures by applying the incremental Isomap and incremental Spati04emporal Isomap (IST-Isomap). Both approaches can produce good real-time video texture results. In particular, IST-Isomap, that we propose, is more suitable for sparse video data (e.g. cartoon)