Learning in Computer Vision and Image Understanding

Using learning in segmentation or recognition tasks has several advantages over classical model-based techniques. These include adaptivity to noise and changing environments, as well as in many cases, a simplified system generation procedure. Yet, learning from examples introduces a new challenge getting a representative data set of examples from which to learn. Applications of learning systems to practical problems have shown that the performance of the system is often critically dependent on both the size and quality of the training set. Federico Girosi of MIT suggested the use of prior information as a general method for synthesizing many training examples from few exemplars. Prototypical transformations are used for general 3D object recognition. Face-recognition was presented as a particular example. Dean Pomerleau of Carnegie Mellon addressed the training data problem as well, within the context of ALVINN, a neural network vision system which drives an autonomous van without human intervention. Some general problems emerge, such as getting sufficient training data for the more unexpected scenes including passing cars and intersections. Several techniques for exploiting prior geometric knowledge during training and testing of the neural-network, were presented. A somewhat different perspective was presented by Bartlett Mel of Caltech. Bartlett introduced a 3D object recognition approach based on concepts from the human visual system. Here the assumption is that a large database of examples exists, with varying viewing angles and distances, as is available to human observers as they manipulate and inspect common objects.