Random Forests with Optimized Leaves for Hough-Voting

Random forest-based Hough-voting techniques are important in numerous computer vision problems such as pose estimation and gesture recognition. Particularly, the voting weights of leaf nodes in random forests have a big impact on performance. We propose to improve Hough-voting with random forests by learning optimized weights of leaf nodes during training. We have investigated two ways for the leaf weight optimization problem by either applying L2 constraints or L0 constraints to those weights. We show that with additional L0 constraints, we are able to simultaneously obtain optimized leaf weights and prune unreliable leaf nodes in the forests, but with additional costs of more computational costs involved during training. We have applied the proposed algorithms to a number of different problems in computer vision, including hand pose estimation, head pose estimation, and hand gesture recognition. The experimental results show that with L2-regularization, regression and classification accuracy are improved considerably. Further, with L0-regularization, many unreliable leaf nodes are suppressed and the tree structure is compressed considerably, while the performance is still comparable to L2-regularization.