Research map / 3D vision / Face and shape
Dr.Hair: Reconstructing Scalp-Connected Hair Strands without Pre-Training via Differentiable Rendering of Line Segments
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Dr.Hair reconstructs scalp-connected hair strands from multi-view images via an optimization pipeline (global 3D orientation estimation, Laplace-equation strand initialization, and a line-segment differentiable rasterizer with Laplacian reparameterization) without any pre-training.
Problem
Reconstructing individual hair strands rooted at the scalp from multi-view surface images typically requires learning-based models pretrained on costly manually created synthetic CG hair data, which introduces domain gaps; prior optimization methods cannot recover internal hair flow with absolute di
Approach
Represent hair as line segments grown from the scalp, estimate globally consistent 3D orientations via MST-based propagation plus a gravity heuristic, fill interior flow by solving Laplace's equation, and optimize strands with a new anti-aliased differentiable rasterizer for line segments plus a Laplacian reparameterization and guide-child hierarchy.
Key results
- On synthetic straight and curly hair at 1mm/10°, Ours reaches F1 52.5 and 29.2 versus NeuralHaircut's 23.0 and 6.6 (curly F1 1mm/10° vs NeuralHaircut 23.2 on curly threshold table)
- Ours at 2mm/20° on curly hair reaches F1 68.9 vs 40.7 for NeuralHaircut
- Speed: surface plus strand reconstruction in under one hour (relative time 25 for strand recon) versus about 2160 for NeuralHaircut, which takes a couple of days
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Dr.Hair instead of NeuralHaircut | multi-view hair strand reconstruction · Straight Hair, 2mm/20° | F1: 42.2 -> 86.2 | Helped |
| Use Dr.Hair instead of NeuralHaircut | multi-view hair strand reconstruction · Curly Hair, 2mm/20° | F1: 23.2 -> 68.9 | Helped |
| Add differentiable-rendering optimization (ablation) | Dr.Hair · Straight Hair, 2mm/20° | F1: 83.4 -> 86.2 | Helped |
| Add differentiable-rendering optimization (ablation) | Dr.Hair · Curly Hair, 2mm/20° | F1: 57.6 -> 68.9 | Helped |
| Optimize guide hair before child hair (ablation) | Dr.Hair · Straight Hair, 2mm/20° | F1: 86.5 -> 86.2 | Hurt |
| Optimize guide hair before child hair (ablation) | Dr.Hair · Curly Hair, 2mm/20° | F1: 68.0 -> 68.9 | Helped |
| Add strand reparameterization (ablation) | Dr.Hair · Straight Hair, 2mm/20° | F1: 38.5 -> 86.2 | Helped |
| Add strand reparameterization (ablation) | Dr.Hair · Curly Hair, 2mm/20° | F1: 37.1 -> 68.9 | Helped |
| Include spatial-proximity neighbors in the reparameterization Laplacian (ablation) | Dr.Hair · Straight Hair, 2mm/20° | F1: 86.0 -> 86.2 | Helped |
| Include spatial-proximity neighbors in the reparameterization Laplacian (ablation) | Dr.Hair · Curly Hair, 2mm/20° | F1: 67.6 -> 68.9 | Helped |
| Initialize strands using gradient-domain strand initialization (ablation) | Dr.Hair · Straight Hair, 2mm/20° | F1: 10.7 -> 86.2 | Helped |
| Initialize strands using gradient-domain strand initialization (ablation) | Dr.Hair · Curly Hair, 2mm/20° | F1: 23.0 -> 68.9 | Helped |
Surprising findings
- Strand initialization alone (w/o DR) already scores well on straight hair, indicating the Laplace-based initialization carries much of the performance there
Limitations
- Output quality depends on the input raw mesh; inaccurate top geometry degrades hair flow
- Smooth-flow assumption cannot handle discontinuous hairstyles such as braids
- Protruding strands disrupt flow, limiting accuracy for highly curly or spiky hair
Experimental setup
Datasets: USC-HairSalon (Yuksel et al. hair models, synthetic test) H3DS
Benchmarks: Yuksel et al. synthetic hair models (58 images) H3DS
Metrics: precision recall F1 runtime
Baselines: LPMVS Strand Integration NeuralHaircut
Builds on: nvdiffrast LPMVS Strand Integration NeuralHaircut Gabor filter orientation estimation
Closest papers by meaning
- MonoHair: High-Fidelity Hair Modeling from a Monocular Video (CVPR 2024 · similarity 0.90)
- DiffLocks: Generating 3D Hair from a Single Image using Diffusion Models (CVPR 2025 · similarity 0.89)
- Doubly Hierarchical Geometric Representations for Strand-based Human Hairstyle Generation (NeurIPS 2024 · similarity 0.89)
- Perm: A Parametric Representation for Multi-Style 3D Hair Modeling (ICLR 2025 · similarity 0.88)
- DGH: Dynamic Gaussian Hair (NeurIPS 2025 · similarity 0.87)
- Text-Conditioned Generative Model of 3D Strand-based Human Hairstyles (CVPR 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.