Abstract

Performance of a machine learning algorithm depends on the representation of the input data. In computer vision problems, histogram based feature representation has significantly improved the classification tasks. L1 normalized histograms can be modelled by Dirichlet and related distributions to transform input space to feature space. We propose a mapping technique that contains prior knowledge about the distribution of the data and increases the discriminative power of the classifiers in supervised learning such as Support Vector Machine (SVM). The mapping technique for proportional data which is based on Dirichlet, Generalized Dirichlet, Beta Liouville, scaled Dirichlet and shifted scaled Dirichlet distributions can be incorporated with traditional kernels to improve the base kernels accuracy. Experimental results show that the proposed technique for proportional data increases accuracy for machine vision tasks such as natural scene recognition, satellite image classification, gender classification, facial expression recognition and human action recognition in videos. In addition, in object tracking, learning parametric features of the target object using Dirichlet and related distributions may help to capture representations invariant to noise. This further motivated our study of such distributions in object tracking. We propose a framework for feature representation on probability simplex for proportional data utilizing the histogram representation of the target object at initial frame. A set of parameter vectors determine the appearance features of the target object in the subsequent frames.

Motivated by the success of distribution based feature mapping for proportional data, we extend this technique for semi-bounded data utilizing inverted Dirichlet, generalized inverted Dirichlet and inverted Beta Liouville distributions. Similar approach is taken into account for count data where Dirichlet multinomial and generalized Dirichlet multinomial distributions are used to map density features with input features.