A classical tool revisited: object detection by statistical testing

Franceschi, F. Odone, F. Smeraldi and A. Verri, in: Proceedings of the International Conference on Instrumentation, Control and Information Technology (SICE), pages 2961 - 2964, Okayama, Japan, August 2005

Abstract

We present a data-driven approach to feature selection and object detection based on hypothesis testing. Starting from positive training examples only, we estimate the probability density of each of a large number of image measurements. A quantitative feature selection criterion inspired by maximum likelihood is then used in conjunction with Spearman's independence rank test to select a maximal subset of discriminative and pairwise independent features. Classification is performed by a sequence of hypothesis tests for the presence of the object. The overall significance level (i.e. the operating point) can be set by controlling the significance level of the individual tests as well as the minimum number of them that a candidate window is required to pass. We report experiments on face detection over the MIT-CBCL database. The image measurements we use for these experiments include grey level values, integral measurements and ranklets. Our results indicate that the method is able to generalize from positive examples only and reaches state-of-the-art recognition rates.

Full paper (PDF)

Backlinks: Publications