Canonical subsets of image features

Abstract

Many object recognition and localization techniques utilize multiple levels of local representations. These local feature representations are common, and one way to improve the efficiency of algorithms that use them is to reduce the size of the local representations. There has been previous work on selecting subsets of image features, but the focus here is on a systematic study of the feature selection problem. We have developed a combinatorial characterization of the feature subset selection problem that leads to a general optimization framework. This framework optimizes multiple objectives and allows the encoding of global constraints. The features selected by this algorithm are able to achieve improved performance on the problem of object localization. We present a dataset of synthetic images, along with ground-truth information, which allows us to precisely measure and compare the performance of feature subset algorithms. Our experiments show that subsets of image features produced by our method, stable bounded canonical sets (SBCS), outperform subsets produced by K-Means clustering, GA, and threshold-based methods for the task of object localization under occlusion.