Maximum likelihood features for generative image models

Abstract

Most approaches to computer vision can be thought of as lying somewhere on a continuum between generative and discriminative. Although each approach has had its successes, recent advances have favored discriminative methods, most notably the convolutional neural network. Still, there is some doubt about whether this approach will scale to a human-level performance given the numbers of samples that are needed to train state-of-the-art systems. Here, we focus on the generative or Bayesian approach, which is more model based and, in theory, more efficient. Challenges include latent-variable modeling, computationally efficient inference, and data modeling. We restrict ourselves to the problem of data modeling, which is possibly the most daunting, and specifically to the generative modeling of image patches. We formulate a new approach, which can be broadly characterized as an application of “conditional modeling,” designed to sidestep the high-dimensionality and complexity of image data. A series of experiments, learning appearance models for faces and parts of faces, illustrates the flexibility and effectiveness of the approach.