Motion Capture Data Completion and Denoising by Singular Value Thresholding

Abstract

Human motion is of high articulation and correlation.When a human motion sequence is represented by a matrix, the matrix will be approximately low-rank. This low-rank property has been used by previous manifold-based approaches such as PCA and GPLVM. Encouraging results yielded by those approaches show that the low-rank property is of interest and importance for animating human motion. However, none of those approaches explicitly exploits it for motion capture data processing. In this paper, we propose to deal with motion capture data based on recently developed low-rank matrix completion theory and algorithms. Unlike previous approaches, the proposed method relies on low-rank prior instead of motion prior. To verify it effectiveness for dealing with motion capture data, we show that incomplete human motion can be effectively reconstructed. We also demonstrate that a noisecorrupted motion can be nicely recovered.