Seminar: Intrinsic Image Decompositions Without Intrinsic Images

Abstract: Modern computer vision relies heavily on very large labelled datasets to  produce complex but effective features that drive very accurate classifiers.  But many important vision problems just don’t yield to this approach.    Intrinsic image decomposition, where one must decompose an image into intrinsic and extrinsic components, is one.  Intrinsic components — like, for example, the fraction of incoming light a surface reflects (its albedo, or lightness) — are scene or object properties that are independent of viewing circumstances.  Extrinsic components — like, for example, the amount of light falling on a surface (its shading) — depend on the viewing circumstances.  A standard problem is to decompose an image into intrinsic and extrinsic components.  In its most usual form, one must recover albedo and shading.  Early methods were unsupervised, relying on spatial reasoning.  More recent methods are supervised, using computer graphics simulations of scenes to train regression networks.  I will describe an unsupervised procedure that uses samples from spatial models (paradigms) to learn a decomposition network.  This procedure is the best known unsupervised method, and outperforms most recent supervised methods. Interestingly, the procedure I will describe makes it possible to recover other intrinsic properties like surface grooves and other extrinsic properties like gloss.  I will describe an application of these methods to image reshading, where one must reshade an image that has been edited to make it look realistic.

Bio: David Forsyth is currently Fulton Watson Copp Chair in Computer Science at U. Illinois at Urbana-Champaign.  He has published papers on computer vision, computer graphics and machine learning. He has served as program co-chair for IEEE Computer Vision and Pattern Recognition in 2000, 2011, 2018, and 2021, general co-chair for CVPR 2006 and 2015 and ICCV 2019, program co-chair for the European Conference on Computer Vision 2008. He has received best paper awards at the International Conference on Computer Vision and at the European Conference on Computer Vision. He received an IEEE technical achievement award for 2005 for his research.  He became an IEEE Fellow in 2009, and an ACM Fellow in 2014.  His textbook, "Computer Vision: A Modern Approach" (joint with J. Ponce and published by Prentice Hall) is now widely adopted as a course text, as is "Probability and Statistics with Computer Science" (Springer), and "Applied Machine Learning" (Springer).   He served two terms as Editor in Chief, IEEE TPAMI, and is a member of a number of Scientific Advisory Boards.