Articulated structure from motion through ellipsoid fitting

Abstract

We present a new method to reconstruct non-rigid objects from orthographic projections based on the assumption of articulated model. We introduce an ellipsoid property to identify points belonging to a rigid subset. This enables us to formulate the problem of motion segmentation as an ellipsoid fitting problem. The obtained rigid subsets are then linked as kinematic chains to constitute 3D articulated structure. This method is practical of computational complexity O(N) mainly based on linear least squares. We demonstrate the effectiveness of this method through experiments on real tracking data, motion capture data, and challenging human dataset with missing data, in comparison with existing methods.