Moving Past Principal Component Analysis: Nonlinear Dimensionality Reduction Towards Better Hand Pose Synthesis

Abstract

Despite their complex kinematic structure with many degrees of freedom, human hands have been shown to have synergistic behavior, with coordinated joint movements being able to explain a large amount of the variance in hand posture measurements. This phenomenon has traditionally been analyzed through Principal Component Analysis (PCA), and has led to important applications in medical robotics, such as the design and control of upper limb prostheses and measurement of hand posture with a reduced number of sensors. However, the use of more complex, nonlinear dimensionality reduction techniques for hand joint measurements has been under-explored in the literature. In this paper, we aim to fill this gap by comparing Principal Component Analysis, Kernel Principal Component Analysis (KPCA), and autoencoders on the same data set, evaluating the performance in terms of Mean Square Error of reconstructed hand poses with respect to the original data set. Results show a better performance for the nonlinear techniques, lowering Mean Square Error up to 25% for the KPCA and 50% for the autoencoders when compared to PCA. Visualization of the reconstructed poses shows a better ability from the autoencoder to reconstruct hand shapes when compared to the two other methods.