Machine Learning for Video-Based Rendering
Arno Schödl, Irfan A. Essa · 2000
We pres[ t techniques for rendering and animation of realis4: ss4: by analyzing and training onsS7] videosoS7(1[(S This workextends the new paradigm for computer animation, video textures, whichus[ recorded video to generate novelanimations by replaying the videos amples in a new order. Here we concentrate on video sprites , which are a s ecial type of video texture. In videos prites ins tead of s oring whole images the object of interes is s eparated from the background and the videos amples are seSF1 as as equence of alpha-matteds prites with ash ciated velocity information. They can be rendered anywhere on the sS een to create a novel animation of the object. We presF t methods to creates uch animations by finding as equence ofs prites amples that is bothvis4 lly sy oth and follows a des7 ed path. To esF[( te visRR s oothnesR we train a linear clas71(S to es imate vis7R s7R] rity between videos amples If the motion pathis known in advance, weus beams earch to find a goods7174 s7174(Sfi We cans pecify the motion interactively by precomputing thesSFR4][ co s functionus[) Q-learning. 1