Predicting and Evaluating the Power of Shared Features

Abstract

Several recent efforts in multi-class feature-based object recognition employ shared features, or features that simultaneously belong to multiple class models. These approaches claim a considerable time savings by reducing the total number of features used by all models, thereby lessening the concomitant computational effort of finding the features in images. In this paper we derive a Bayesian framework for predicting and evaluating the performance of shared feature-based recognition systems. We then use this framework to predict the performance of several instances of a simple multi-class object detector.