High-Quality Hair Modeling from A Single Portrait Photo

摘要

We propose a novel system to reconstruct a high-quality hair depthrnmap from a single portrait photo with minimal user input. Wernachieve this by combining depth cues such as occlusions, silhouettes,rnand shading, with a novel 3D helical structural prior forrnhair reconstruction. We fit a parametric morphable face model tornthe input photo and construct a base shape in the face, hair andrnbody regions using occlusion and silhouette constraints. We thenrnestimate the normals in the hair region via a Shape-from-Shadingbasedrnoptimization that uses the lighting inferred from the facernmodel and enforces an adaptive albedo prior that models the typicalrncolor and occlusion variations of hair. We introduce a 3D helicalrnhair prior that captures the geometric structure of hair, and showrnthat it can be robustly recovered from the input photo in an automaticrnmanner. Our system combines the base shape, the normalsrnestimated by Shape from Shading, and the 3D helical hair priorrnto reconstruct high-quality 3D hair models. Our single-image reconstructionrnclosely matches the results of a state-of-the-art multiviewrnstereo applied on a multi-view stereo dataset. Our techniquerncan reconstruct a wide variety of hairstyles ranging from short tornlong and from straight to messy, and we demonstrate the use ofrnour 3D hair models for high-quality portrait relighting, novel viewrnsynthesis and 3D-printed portrait reliefs.