Title: Towards High Quality Hair Capture and Modeling
Abstract:
Hair is one of human's most distinctive features and one important
component in human digitalization. However, capturing and modeling
hair remains challenging because of hair's unique properties: the
specular appearance violates the Lambertian surface assumption used in
most multi-view stereo reconstruction methods; The vast variations and
complex topology of hair styles pose great technical difficulties for
high quality hair capture and modeling.
We propose an orientation-based matching metric for multi-view hair
capture that is more robust to hair's view-dependent specular
highlights. The metric is performed in multi-resolution to reveal
increasing hair structures and details. The orientation-based metric
also facilitates the structure-aware aggregation to reduce matching
error and stereo noise along coherent hair structures. Our approach is
able to reconstruct approximate surfaces to a variety of hair styles
with detailed hair structures from ~30 input views.
To alleviate the requirement for dense small-baseline capture setup,
we introduce a visual hull based refinement method to enable
wide-baseline full hair capture with only 8 views. The refinement is
driven by the strands generated from the orientation map of each input
view. The final shape is obtained by optimizing the orientation
consistency of these strands against all the orientation maps as well
as the regularization terms that account for the smoothness at strand,
wisp and global levels. We find our method effective on various hair
styles and suitable for dynamic hair capture. We achieve an
average reconstruction accuracy of ~3mm on synthetic datasets.
Finally, we address the challenge of robustly modeling hair strands
from the captured incomplete exterior hair geometry with plausible
topology for hair simulation. Our main contribution is the
construction and analysis of the global connection graph to compute
the plausible hair topology from the incomplete exterior hair
geometry. The connection graph encodes coherent local strands and
plausible inter-strand connections. The final hair strands can be
generated from a wisp forest derived from the connection graph.