Abstract
Robust segmentation of hair from portrait images remains challenging: hairdoes not conform to a uniform shape, style or even color; dark hair inparticular lacks features. We present a novel computational imaging solutionthat tackles the problem from both input and processing fronts. We exploreusing Time-of-Flight (ToF) RGBD sensors on recent mobile devices. We firstconduct a comprehensive analysis to show that scattering and inter-reflectioncause different noise patterns on hair vs. non-hair regions on ToF images, bychanging the light path and/or combining multiple paths. We then develop a deepnetwork based approach that employs both ToF depth map and the RGB gradientmaps to produce an initial hair segmentation with labeled hair components. Wethen refine the result by imposing ToF noise prior under the conditional randomfield. We collect the first ToF RGBD hair dataset with 20k+ head imagescaptured on 30 human subjects with a variety of hairstyles at different viewangles. Comprehensive experiments show that our approach outperforms the RGBbased techniques in accuracy and robustness and can handle traditionallychallenging cases such as dark hair, similar hair/background, similarhair/foreground, etc.