Full Head Performance Capture Using Multi-scale Mesh Propagation.

We present a template fitting based method for efficient markerless full head performance capture. Our method starts with high-resolution multi-view videos and efficiently outputs high-fidelity full head mesh sequences with the same topology of a common template model. A GPU accelerated stereo reconstruction firstly computes a high-quality point cloud at each frame. Then the template model is warped and fitted to the reconstructed geometries using a combination of detected landmarks constraint, nearest neighbor constraint, and volumetric regularization. Additionally, we reconstruct the detailed ear structures at the initialization frame and track the movement with a global rigid transformation assumption in the following frames. To solve the error accumulation problem when dealing with long sequences, the method updates the positions of mesh vertices using priors from the previous and initial frames in a coarse to fine manner. Summing up the above technical innovations, our method can significantly reduce the whole processing time for reconstructing topology consistent full head meshes with fine details. We conduct several experiments and demonstrate the efficiency and out-performance of the proposed method compared to previous methods.