Hair is one of the most challenging objects to reconstruct due to its micro-scale structure and a large number of repeated strands with heavy occlusions. We present the first method to capture high-fidelity hair geometry with strand-level accuracy. Our approach consists of three stages. First, we develop a novel multi-view stereo technique with slanted support lines tailored for hair, which capture the hair correspondence across views. Second, we propose a strand reconstruction method with a mean-shift algorithm that converts the captured point cloud into distinct strands. Third, we use multi-view geometric constraints to elongate incomplete strands and recover undetected ones. We validate our method using both synthetic and real-world captured data and demonstrate sub-millimeter accuracy in reconstructed hair geometry.
①
Novel multi-view stereo with slanted support lines to establish accurate hair correspondences across camera views, producing a dense 3D point cloud.
②
Mean-shift algorithm converts the captured 3D point cloud into coherent and distinct hair strands with accurate orientations.
③
Multi-view geometric constraints elongate incomplete strands and recover missing ones, achieving sub-millimeter overall accuracy.
@inproceedings{nam2019strand,
title={Strand-accurate Multi-view Hair Capture},
author={Nam, Giljoo and Wu, Chenglei and Kim, Min H. and Sheikh, Yaser},
booktitle={IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
year={2019}
}