Professor David A. Forsyth from the University of Illinois at Urbana-Champaign, gave a lecture on November 23, 2017 titled “New Scenes by Decomposition and Reassembly”.
Professor Forsyth described methods that can decompose pictures and reassemble them into pictures of new scenes. First, he described methods to insert computer graphics objects into images of real scenes, while adjusting lighting etc. so as to make it look as if the object "belongs" in the picture. Second, he described methods to cut objects out of one picture and insert them into another. This process is more complicated, because one must infer a shape representation for the image fragment. Third, he described methods that reassemble pictures from a decomposition into useful layers. The methods rely on novel conditional image synthesis architectures that can learn to produce multiple, diverse samples of a particular layer consistent with a given input. For example, one might recover the shading of an image, throw that away, then synthesize and apply new shading layers that are consistent with the original albedo, giving the impression the illumination field has changed. Fourth, He showed how the methods can be adapted to edit faces in images, changing makeup, lighting, and some complexion and appearance details. Finally, Prof. Forsyth suggested new directions of research in image synthesis network architectures to produce better, more useful, image editing systems.