Discriminative Dictionary Design for Action Classification in Still Images

Abstract

In this paper, we address the problem of action recognition from still images. Although used widespread, local features (SIFT, STIP) invariably engender two potential problems: the counts of such extracted features are not evenly distributed in different entities of a given category and many of such features are not paradigmatic of the visual concept the entities represent. In order to generate a discriminative dictionary taking the aforementioned issues into account, we propose a novel method for identifying robust and category specific local features which maximize the class separability to a possible extent. Specifically, we consider category independent region proposals to highlight local regions in still images. Further, the selection of potent local descriptors is cast as filtering based feature selection problem which ranks the local features per category based on a novel measure of distinctiveness. The underlying visual entities are subsequently represented based on the learned dictionary and this stage is followed by action classification using the random forest model. The framework is validated on the challenging Stanford-40 dataset and exhibits superior performances than the representative methods from the literature.