How HairStep Revolutionizes Single-View 3D Hair Reconstruction

Abstract

With the growing use of virtual reality in entertainment, education, and social domains, high‑quality 3D hair models are increasingly demanded. Existing pipelines rely on costly 3D hair designs by artists. This work proposes reconstructing high‑quality 3D hair from a single portrait image using a new intermediate representation called HairStep , which consists of a Strand Map and a Depth Map . HairStep reduces the domain gap between synthetic training data and real images.

HairStep Representation

Given an RGB portrait, HairStep encodes hair geometry as follows:

Strand Map : the red channel stores a hair mask, while the green and blue channels store normalized 2‑D hair strand directions.

Depth Map : each pixel records the distance from the camera to the nearest 3‑D point on the hair surface.

Dataset Construction

Two datasets were built:

HiSa : 1,250 portrait images annotated with dense 2‑D hair curves (≈300 per image). These curves are rasterized into a Stroke Map and interpolated with the hair mask to produce the Strand Map.

HiDa : a relative depth dataset where paired pixel points in hair regions are annotated with ordinal depth values, providing weak supervision for depth estimation.

Extraction of HairStep

Strand Map is predicted from RGB images using a U‑Net trained with L1 and perceptual losses on the HiSa dataset. Compared with orientation maps generated by Gabor filtering, the learned Strand Map exhibits less noise and clearer hair direction.

Depth Map is obtained via a domain‑adaptive depth estimation network based on an Hourglass architecture. The network is first pretrained on synthetic data, then refined using pseudo‑labels generated from the Ground‑Truth Strand Map of HiSa and the ordinal depth pairs from HiDa.

3D Hair Reconstruction Experiments

Three state‑of‑the‑art single‑view 3D hair reconstruction methods were evaluated:

HairNet (explicit point sequence)

DynamicHair (voxel orientation field)

NeuralHDHair (implicit orientation field)

Using HairStep as input consistently improved reconstruction accuracy for all three algorithms, reducing orientation error by up to 50.3% compared with traditional orientation maps.

Evaluation Metrics

A fair quantitative evaluation was introduced by rendering the reconstructed 3D hair back to Strand and Depth Maps and comparing them with ground‑truth annotations. Two metrics were defined:

HairSale : average angular error of 2‑D hair strand directions.

HairRida : accuracy of relative depth ordering.

Conclusion and Outlook

HairStep provides a rich intermediate representation that enables accurate single‑view 3D hair reconstruction across diverse algorithms. The released HiSa and HiDa datasets, along with the new evaluation metrics, offer valuable resources for future research. Limitations include reduced performance on braids and highly complex curls.