Hierarchical Metric Learning for Fine Grained Image Classification
share
This paper deals with the problem of fine-grained image classification and introduces the notion of hierarchical metric learning for the same. It is indeed challenging to categorize fine-grained image classes merely in terms of a single level classifier given the subtle inter-class visual differences. In order to tackle this problem, we propose a two stage framework where i) the image categories are represented hierarchically in terms of a binary tree structure where different subset of classes are present in the non-leaf nodes of the tree. This is accomplished in an automatic fashion considering the available training data in the visual domain, and ii) a (non-leaf) node specific metric learning is further deployed for the categories constituting a given node, thus enforcing better separation between both of its children. Subsequently, we construct (non-leaf) node specific binary classifiers in the learned metric spaces on which testing is henceforth carried out by following the outcomes of the classifiers sequence from root to leaf nodes of the tree. By separately focusing on the semantically similar classes at different levels of the hierarchy, it is expected that the classifiers in the learned metric spaces possess better discriminative capabilities than considering all the classes at a single go. Experimental results obtained on two challenging datasets (Oxford Flowers and Leeds Butterfly) establish the superiority of the proposed framework in comparison to the standard single metric learning based methods convincingly
READ FULL TEXT