Markerless Motion Capture: the goals, the vision and (reaching) the impossible

scenarios without any markers on the subject, a set-up can become arbitrarily complex. In the talk we outline some
achievements of our markerless motion capture system we developed in the last three years. The input is a multi-view
image stream and a representation of the subject in terms of free-form surface patches. Our system extracts silhouettes
using level set functions, determines correspondences between the model and the image data and finally computes the pose
configuration (the rotation and translation in 3D and the joint angles).

The talk further presents extensions of this basic set-up, e.g. tracking clothed people, high-accuracy tracking (our
performance in the HumanEVA benchmark), the integration of motion priors and statistical learning, tracking textured
object models or constricted kinematic chains. Several example video sequences document the improvements and advances.