Efficient Pose Estimation using Random Forest and Hash Voting

Pose estimation is one of the key components in robot perception and exhibits a number of unique challenges. First, it is non-trial to directly search for potential poses in given images. Second, it is very challenging to retrieve pose features hidden in images or point clouds in the presence of textureless objects and occlusion. We present a pose estimation pipeline using RGBD images. We first use random forest to perform segmentation and locate the object of interest in a given RGBD image. Then we generate sufficient hypotheses and compute their possibility distribution using hash voting. Our results show high precision and good performance under severe conditions: textureless objects and occlusion.