The goal of the proposed research is to develop and test a computational theory for the human ability to recognize objects under variable illumination (including extreme shadowing) and viewpoint changes. The ability to recognize objects is of fundamental importance in everyday life and the loss of this ability, due to a stroke or Alzheimer's disease, is a serious handicap to the person involved. The computational theory is based on a new paradigm for object representation --generative modeling - - in which an image-based model of an object is """"""""generated"""""""" from a small set of training images. This theory has been demonstrated to successfully recognize objects from real images under extreme lighting variations. This gives a reality check on the theory and can be thought of as making it an ecological theory (in the sense that it yields good results on the types of images that humans encounter in the real world and not just on the visual stimuli occurring in laboratories). We have assembled a team of researchers with interdisciplinary skills in computer and biological vision. who will divide their efforts on the project based on their expertise. It is our explicit intent that the algorithms and psychophysical studies develop in tandem, with each group verifying the other's results. Indeed, as reviewed below, the computer vision theory, when applied to human performance, makes a number of predictions. some of which have already been partially confirmed by our preliminary experimental work. Our proposal is organized into three main areas. The psychophysical work parallels the computational issues in three series of experiments in which we investigate: (I) How human observers learn and recognize objects, given variable lighting conditions, from a single fixed viewpoint. (II) How illumination and viewpoint interact in human object recognition. (III) The role of class-specific knowledge in recognition.
| Lee, Kuang-Chih; Ho, Jeffrey; Kriegman, David J (2005) Acquiring linear subspaces for face recognition under variable lighting. IEEE Trans Pattern Anal Mach Intell 27:684-98 |
| Kersten, Daniel; Mamassian, Pascal; Yuille, Alan (2004) Object perception as Bayesian inference. Annu Rev Psychol 55:271-304 |
| Kersten, Daniel; Yuille, Alan (2003) Bayesian models of object perception. Curr Opin Neurobiol 13:150-8 |
| Naor-Raz, Galit; Tarr, Michael J; Kersten, Daniel (2003) Is color an intrinsic property of object representation? Perception 32:667-80 |
| Yuille, A L; Rangarajan, Anand (2003) The concave-convex procedure. Neural Comput 15:915-36 |
| Yuille, Alan; Coughlan, James M; Konishi, Scott (2003) The KGBR viewpoint-lighting ambiguity. J Opt Soc Am A Opt Image Sci Vis 20:24-31 |
| Yuille, A L (2002) CCCP algorithms to minimize the Bethe and Kikuchi free energies: convergent alternatives to belief propagation. Neural Comput 14:1691-722 |
