Hierarchical spectral clustering based large margin classification of visually correlated categories

Object recognition is one of the challenging tasks in computer vision and the problem becomes increasingly difficult when the image categories are visually correlated among themselves i.e. they are visually similar and only fine differences exist among the categories. This paper has a two-fold objective which involves organization of the image categories in a hierarchical tree like structure using self tuning spectral clustering for exploiting the correlations among them. The organization phase is followed by a node specific large margin nearest neighbor classification scheme, where a Mahalnobis distance metric is learnt for each non-leaf node. Further a procedure for hyperparameters selection has been discussed w.r.t two strategies i.e. grid search and Bayesian optimization. The proposed algorithm's effectiveness is tested on selected classes of the popular Imagenet dataset.