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

**Authors:** Yusuke Takimoto, Hikari Takehara, Hiroyuki Sato, Zihao Zhu, Bo Zheng  
**arXiv:** [2403.17496](https://arxiv.org/abs/2403.17496)  
**Date:** 2024-03-26  
**Summary page:** https://www.summarizepaper.com/en/paper/2403.17496v2/

## Key points

- Achieving a realistic hair appearance is crucial in film and gaming industries for creating immersive experiences.
- Traditional methods involve using strands from the scalp to mimic natural hair growth patterns, but this poses challenges due to the complexity of real-world hair textures.
- State-of-the-art learning-based methods have relied on pre-training with synthetic computer-generated data, which is labor-intensive and costly.
- The researchers' method represents hair strands as line segments growing from the scalp and optimizes them using a differentiable rendering algorithm.
- This approach eliminates the need for pre-training by leveraging 3D orientation estimation, strand initialization based on Laplace's equation, and reparameterization considering geometric connectivity and spatial proximity.
- The new method excels at reconstructing internal hair flow directionality in an absolute direction, surpassing existing optimization-based methods.
- Through testing, the researchers demonstrated robust and accurate inverse rendering capabilities that improve processing speed while maintaining reconstruction accuracy.

## Summary

In the film and gaming industries, achieving a realistic hair appearance is crucial for creating immersive experiences. Traditionally, this has involved the use of strands originating from the scalp to mimic natural hair growth patterns. However, reconstructing these strands accurately from observed surface images of hair poses significant challenges due to the complexity and variability of real-world hair textures. To address this issue, state-of-the-art learning-based methods have relied on pre-training with manually prepared synthetic computer-generated (CG) data. While effective, this approach is labor-intensive, costly, and introduces complications due to the domain gap between synthetic and real-world data. The researchers' method represents hair strands as line segments growing from the scalp and optimizes them using a novel differentiable rendering algorithm. This approach eliminates the need for pre-training by leveraging 3D orientation estimation through global optimization, strand initialization based on Laplace's equation, and reparameterization that considers geometric connectivity and spatial proximity. Unlike existing optimization-based methods that struggle with internal hair flow directionality, this new method excels at reconstructing internal hair flow in an absolute direction. Through rigorous testing and evaluation, the researchers demonstrated that their method exhibits robust and accurate inverse rendering capabilities. Not only does it surpass the quality of existing methods in terms of reconstruction accuracy but it also significantly improves processing speed. By eliminating the reliance on pre-training with synthetic data and offering superior reconstruction results without sacrificing efficiency,this novel approach presents a promising advancement in realistic hair modeling for film and gaming applications.