Machine Learning for Video-Based Rendering
Essa, Irfan A.
MetadataShow full item record
We recently introduced a new paradigm for computer animation, video textures, which allows us to use a recorded video to generate novel animations by replaying the video samples in a new order. Video sprites are a special type of video texture. Instead of storing whole images, the object of interest is separated from the background and the video samples are stored as a sequence of alpha-matted sprites with associated velocity information. They can be rendered anywhere on the screen to create a novel animation of the object. To create such an animation, we have to find a sequence of sprite samples that is both visually smooth and shows the desired motion. In this paper, we address both problems. To estimate visual smoothness, we train a linear classifier to estimate visual similarity between video samples. If the motion path is known in advance, we then use a beam search algorithm to find a good sample sequence. We can also specify the motion interactively by precomputing a set of cost functions using Q-learning.