Learning Statistical Structure for Object Detection

Abstract

Many classes of images exhibit sparse structuring of statistical dependency. Each variable has strong statistical dependency with a small number of other variables and negligible dependency with the remaining ones. Such structuring makes it possible to construct a powerful classifier by only representing the stronger dependencies among the variables. In particular, a semi-na?e Bayes classifier compactly represents sparseness. A semi-na?e Bayes classifier decomposes the input variables into subsets and represents statistical dependency within each subset, while treating the subsets as statistically inde-pendent. However, learning the structure of a semi-na?e Bayes classifier is known to be NP complete. The high dimensionality of images makes statistical structure learning especially challenging. This paper describes an algorithm that searches for the structure of a semi-na?e Bayes classifier in this large space of possible structures. The algorithm seeks to optimize two cost functions: a localized error in the log-likelihood ratio function to restrict the structure and a global classification error to choose the final structure. We use this approach to train detectors for several objects including faces, eyes, ears, telephones, push-carts, and door-handles. These detectors perform robustly with a high detection rate and low false alarm rate in unconstrained settings over a wide range of variation in background scenery and lighting.