Hand Coordinate Regression (Minor Thesis)

Abstract: This thesis explores a novel approach for one-shot hand pose estimation from a single RGB-D image pair. The idea is to use a random forest to predict dense features as an intermediate regression step, which can subsequently be optimized to obtain a final pose estimate. This work addresses the design of the intermediate representation by presenting several concepts for its definition in form of \textit{hand coordinates}. Random forests are then trained to regress these hand coordinates and are tested for each concept. The results show that the obtained dense features are suitable for further refinement towards a final pose estimate; however, since they are an intermediate step in the pose estimation process, a final assessment of the quality of the presented approaches cannot be made yet. They do look promising as a starting point for further optimization though. As the random forests required a lot of training and test data, this thesis also contributes new ground truth to an existing hand pose data set for each of the presented hand coordinate concepts.