By Peter Carbonetto (joint work with Nando de Freitas)
In my current work with Nando de Freitas, I'm experimenting with various methods to construct Bayesian models for labeling objects in a scene. The model is provided with a training set (images labeled with words), and the objective is to generalize this to unseen examples as optimally as possible. In the talk I will point out some of the challenges in building such a image translation model, and how we propose to surmount them.
Check out these links for more information on our object recognition model and our for feature weighting.