Training Deformable Models for Localization in Images

Deva Ramanan

Toyota Technical Institute

We present a new method for training deformable models. Assume that we have training images of objects (say, a horse) where parts have been labeled. Typically, one fits a model by maximizing the likelihood of the part labels. Alternatively, one could fit a model such that, when the model is run on the training images, it finds the parts. We show this criteria leads to very different models because the objective is discriminative (rather than generative) and model parameters are jointly learned (rather than independently). We formulate model-learning as parameter estimation in a conditional random field (CRF); this means we can learn globally optimal parameters by simple gradient ascent. We present results on three established vision datasets of motorcycles, people, and horses that achieve or surpass the state-of-the-art. This is joint work with Cristian Sminchesescu.

Deepak Ramachandran

Last modified: Thu Jan 19 14:25:50 CST 2006