Learning Marginalization through Regression for Hand Orientation Inference

We present a novel marginalization method for multilayered Random Forest based hand orientation regression. The proposed model is composed of two layers, where the first layer consists of a marginalization weights regressor while the second layer contains expert regressors trained on subsets of our hand orientation dataset. We use a latent variable space to divide our dataset into subsets. Each expert regressor gives a posterior probability for assigning a given latent variable to the input data. Our main contribution comes from the regression based marginalization of these posterior probabilities. We use a Kullback-Leibler divergence based optimization for estimating the weights that are used to train our marginalization weights regressor. In comparison to the state-of-the-art of both hand orientation inference and multi-layered Random Forest marginalization, our proposed method proves to be more robust.