by Matthew Brown
This talk will describe a new approach to alpha matting using Markov Random Fields. This enables users to cut and paste objects in images by simply specifying areas of texture, instead of laboriously tracing a contour. Recent approaches to alpha matting have used learning and inference with probabilistic models of colour. In this work we extend this idea by learning models of complete textures and using these for inference of alpha mattes.
This is joint work with Andrew Blake at Microsoft Research, Cambridge, UK.