Learning balanced trees for large scale image classification

Abstract

The label tree is one of the popular approaches for the problem of large scale multi-class image classification in which the number of class labels is large, for example, several tens of thousands of labels. In learning stage, class labels are organized into a hierarchical tree, in which each node is associated with a subset of class labels and a classifier that determines which branch to follow; and each leaf node is associated with a single class label. In testing stage, the fact that a test example travels from the root of the tree to a leaf node reduces the test time significantly compared to the approach of using multiple binary one-versus-all classifiers. The balance of the learned tree structure is the key essential of the label tree approach. Previous methods for learning the tree structure use clustering techniques such as k-means or spectral clustering to group confused labels into clusters associated with the nodes. However, the output tree might not be balanced. We propose a method for learning effective and balanced tree structure by jointly optimizing the balance constraint and the confusion constraint. The experimental results on the datasets such as Caltech-256, SUN-397, and ImageNet-1K show that the classification accuracy of the proposed approach outperforms that of other state of the art methods.