Dr.Hair: Reconstructing Scalp-Connected Hair Strands without Pre-training via Differentiable Rendering of Line Segments

TL;DR

Dr.Hair presents a pre-training-free, optimization-based pipeline for reconstructing scalp-connected hair strands from multi-view images using differentiable rendering of line segments. It introduces a global optimization step to produce consistent 3D orientations, a Laplace-based interior-strand initialization to model internal hair flow, and a reparameterized, Laplacian-guided optimization with a guide–child strand hierarchy. The method avoids domain gaps from synthetic priors, achieves robust performance on both synthetic and real data, and delivers faster processing (often under an hour) compared with prior approaches that require extensive pre-training. By combining traditional scalp modeling, gradient-domain initialization, and a differentiable renderer for line segments, Dr.Hair yields accurate hair direction, dense strand reconstruction, and realistic rendering suitable for photorealistic character digitization and animation.

Abstract

In the film and gaming industries, achieving a realistic hair appearance typically involves the use of strands originating from the scalp. However, reconstructing these strands from observed surface images of hair presents significant challenges. The difficulty in acquiring Ground Truth (GT) data has led state-of-the-art learning-based methods to rely on pre-training with manually prepared synthetic CG data. This process is not only labor-intensive and costly but also introduces complications due to the domain gap when compared to real-world data. In this study, we propose an optimization-based approach that eliminates the need for pre-training. Our method represents hair strands as line segments growing from the scalp and optimizes them using a novel differentiable rendering algorithm. To robustly optimize a substantial number of slender explicit geometries, we introduce 3D orientation estimation utilizing global optimization, strand initialization based on Laplace's equation, and reparameterization that leverages geometric connectivity and spatial proximity. Unlike existing optimization-based methods, our method is capable of reconstructing internal hair flow in an absolute direction. Our method exhibits robust and accurate inverse rendering, surpassing the quality of existing methods and significantly improving processing speed.

Figures (35)

• Figure 1: Results of existing strand-based 3D reconstruction methods and our method tested with the data captured by a multi-camera system. In the upper row, color and colored arrows represent 3D orientation of hair strands. The overlaid black arrows were drawn manually to visualize rough orientations. The lower row shows individual strands with random color. LPMVS and Strand Integration failed to estimate consistent direction, and their strands are too short not to connect to the scalp. The absolute orientation of strands estimated by NeuralHaircut is mostly the opposite of the actual hair orientation. Our method demonstrates better precision in reconstructing the directional flow of scalp-connected hair.
• Figure 2: The overview of our pipeline. Our approach combines traditional real-time rendering techniques with recent advances in differentiable rendering. First, we fit a template to a raw mesh. Next, we compute consistent 3D orientations from 2D orientation images and initialize guide strands based on a differential equation. Finally, optimization based on differentiable rendering is applied by leveraging the hierarchical relationship between guides and children.
• Figure 3: Raw mesh
• Figure 4: Before optim.
• Figure 5: MST