Semisupervised online learning of hierarchical structures for visual object classification

One of the main challenges in hierarchical object classification is the derivation of the correct hierarchical structure. The classic way around the problem is assuming prior knowledge about the hierarchical structure itself. Two major drawbacks result from the former assumption. Firstly it has been shown that the hierarchies tend to reduce the differences between adjacent nodes. It has been observed that this trait of hierarchical models results in a less accurate classification. Secondly the mere assumption of prior knowledge about the form of the hierarchy requires an extra amount of information about the dataset that in many real world scenarios may not be available. In this work we address the mentioned problems by introducing online learning of hierarchical models. Our models start from a crude guess of the hierarchy and proceed to figure out the detailed version progressively. We show the merits of the proposed work via extensive simulations and experiments on a real objects database.