Constructing Hierarchical Visual Tree for Discriminative Image Representation and Classification

To support large-scale visual recognition, it is critical to train a large number of classifiers with high discrimination power. To achieve this task, in this paper a hierarchical visual tree is constructed for organizing a large number of object classes and image concepts according to their inter-concept visual correlations. Based on the hierarchical visual tree, a novel approach is proposed for learning multi-scale group-based dictionary to support discriminative bag-of-visual-words (BoW-based) image representation. In addition, a structural learning approach is developed to enable large-scale classifier training over such hierarchical visual tree. We have also compared the performance of our hierarchical visual tree with traditional label tree over large-scale image collections.