Dynamic Hair Capture

Abstract:

The realistic reconstruction of hair motion is challenging because of
hair's complex occlusion, lack of a well-defined surface, and
non-Lambertian material. We present a system for passive capture of
dynamic hair
performances using a set of high-speed video cameras. Our key
insight is that, while hair color is unlikely to match across multiple
views, the response to oriented filters will. We combine a multi-scale
version of this orientation-based matching metric with bilateral
aggregation, a MRF-based stereo reconstruction technique, and algorithms
for temporal tracking and de-noising. Our final output is a set of hair
strands for each frame, grown according to the per-frame reconstructed
rough geometry and orientation field. We demonstrate results for a
number of hair styles ranging from smooth and ordered to curly and
messy.