Learning compact and efficient deformable models of human shape variation

Our main proposed contribution is to investigate innovative group-wise approaches to learning compact and efficient deformable models of the human shape variation, and to examine their effectiveness in improving the precision of markerless motion capture systems. Of particular novel interest is the use of such approaches to assist in the kinematic analysis of individuals participating in sporting activities.
The problem of efficiently recovering the pose (3D position and orientation) of human body parts using visual markerless observations is an interesting problem, due to its theoretical importance and its potential diverse uses. It is a problem which the human visual system exhibits a remarkable ability to seemingly effortlessly solve this complex problem, however one which remains an extremely difficult task for a computer to replicate with the same level of accuracy.

A wide range of useful applications can be implemented provided that this fundamental problem is robustly and efficiently solved, in particular motion capture systems. Traditionally, they employ optical markers and / or other specialized hardware to tackle this problem and are widely used especially in the entertainment industry. However, a large body of literature exists which is devoted to real-time recovery of pose for markerless articulable objects, such as human bodies, clothes, and man-made objects. The high level of interest in developing markerless computer-vision based solutions is due to the fact that they are non-invasive, more flexible and potentially cheaper than solutions based on intrusive optical marker based systems.

In order to improve the performance of these solutions, learned generic morphable models which capture human shape variation, including complex non-rigid deformations and articulation, have been shown to be particularly useful. A common approach attempts to parametrization the entire shape space represented by the data in the form of a lower dimensional subspace, in which the variability of an object class from a template or mean shape is encoded.

Despite a number of drawbacks, optical-marker based techniques remain the mainstay for motion capture for the film and video-game industries. This is due to most current approaches being imprecise, limited to working with single individuals and unable to cope when there are interactions with peripheral items in the scene, such as props or sporting equipment. For future markerless systems to be more successful and less dependent on various assumptions, learning highly detailed shape models provides a credible research direction.