Articulated tree structure from motion—A matrix factorisation approach

We present a matrix factorisation approach for 3D reconstruction of articulated objects from monocular images. Previous factorisation methods can solve for multiple independently moving objects, but fail in the articulated case because of a rank deficiency in the input matrix. By extending the shape and motion matrices used in the case of independent objects, we derive a reconstruction algorithm for articulated motion. The recovered 3D shape includes information about the structure of each articulated segment and the point of rotation of each segment relative to its parent segment. Preliminary experiments on synthetic data indicate that noise tolerance compares favourably to that reported for the case of independently moving objects. We also present results quantifying the effect of input noise on the reconstruction accuracy.