Motion Interpretation and Synthesis by ICA

It is known that high-dimensional human motion data lies in a low dimensional space. Looking for independent sub-motions that contribute to generating complicated human motions is an interesting task on its own merits. Independent Component Analysis (ICA) can extract independent factors of an observed signal. In this paper, ICA is applied to the reconstructed human motion data by PCA. Experimental results show that, if the low dimenality is appropriately chosen, ICA can find meaningful sub-motions. Independent sub-motions make it easy to complete missing human motion data. Preliminary results on synthesizing missing human motion data by self-regression using Gassian Process will also be discussed in the paper.