Easy minimax estimation with random forests for human pose estimation

Abstract

We describe a method for human parsing that is straightforward and competes with state-of-the-art performance on standard datasets. Unlike the state-of-the-art, our method does not search for individual body parts or poselets. Instead, a regression forest is used to predict a body configuration in body-space. The output of this regression forest is then combined in a novel way. Instead of averaging the output of each tree in the forest we use minimax to calculate optimal weights for the trees. This optimal weighting improves performance on rare poses and improves the generalization of our method to different datasets. Our paper demonstrates the unique advantage of random forest representations: minimax estimation is straightforward with no significant retraining burden.