A deeper look at hand pose estimation

Abstract

Hand pose recognition is a fundamental human ability and an important, yet elusive, goal for computer vision research. One of the major challenges in hand pose recognition is the sheer scale of the problem. The human hand is a notoriously agile object with 27 degrees of freedom. In a sense, it is an impossible task to collect a dataset with every major hand pose configuration. However, current state-of-the-art approaches rely too much on training data and generalize poorly to unseen hand poses. Furthermore, current benchmarking datasets are of poor quality and contain test sets that are highly correlated with the training set, which in turn encourages the development of data-reliant techniques for better accuracy only on paper. In this thesis, I introduce a better and more realistic benchmarking dataset, and propose a novel approach for hand pose detection that has the potential to generalize better to unseen hand poses.