Articulated structure from motion by factorization

Multibody affine Structure From Motion (SFM) methods commonly assume independent motion between objects such that the `measurement matrix' has rank 4k. When multiple views are available, each object is then independently calibrated to a metric co-ordinate frame. However, articulated motion results in a further decrease in rank - a fact that we exploit to detect articulated objects and determine their degrees of freedom using simple linear methods. Furthermore, these objects cannot be recovered and calibrated independently since this violates articulation constraints. We show that articulation constraints can be imposed during factorization and self-calibration to recover consistent 3D structure and motion, from which link lengths and joint angles can be computed. The stability of the method is evaluated using synthetic data for comparison with ground truth and results are also presented for real image sequences.