Three-Dimensional Hair Structure Reconstruction from a Single Sketch Image without Intermediate Representation

Abstract

Reconstructing 3D hair structures from a single image is a highly promissing research direction. However, current methods only focus on real-world images, while user-oriented applications require more freedom in input data. Existing methods for buildings 3D hair from sketchy images use synthetic data for training, which encounters the domain gap issue. We experiment with building a dataset directly from hand-drawn sketches and propose a model trained on it. As a result, without using an intermediate oriented map representation, the model is still able to learn how to reconstruct hair at a satisfactory level. This opens up a new direction for this problem.