We investigate a free-form inference of the 3D information about the world that can be extracted directly from image sequences taken from a moving camera. Such inference can only succeed if certain assumptions are made, the standard being that the scene observed by the camera is rigid.
Recent studies on non-rigid factorisation have demonstrated that it is possible, under certain viewing conditions, to infer the principal modes of deformation of an object alongside its 3D shape within a structure from motion estimation framework. The models recovered by these algorithms, in which deformations are represented as linear combinations of a number of detected modes, can subsequently be used as compact representation of the object suitable for use in tracking, animation, or other analysis.
We are aiming both to extend and improve factorisation methods and to apply them to the domain of human motion analysis - in particular to the 3-D reconstruction of facial motion.