Research map / 3D vision / 3D generation
MonoHair: High-Fidelity Hair Modeling from a Monocular Video
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.
Reconstructs strand-based 3D hair from a monocular video by refining a NeRF-derived raw point cloud into an exterior line map via patch-based multi-view optimization, then inferring interior structure with an improved DeepMVSHair trained on synthetic data.
Problem
High-fidelity strand-level 3D hair reconstruction from casually captured monocular video, without strict capture conditions and without over-relying on learned hair priors that smooth out fine details such as curly hair.
Approach
Split hair modeling into exterior and interior: the exterior is reconstructed prior-free from multi-view image cues (patch-wise angular loss plus ray regularization) and then rendered as undirectional strand maps and depth maps that feed a data-driven interior inference network, closing the synthetic-to-real domain gap.
Key results
- On the synthetic benchmark at 2/20 thresholds, the method reaches 60.8 precision and 17.8 F-score, versus 52.9 and 16.4 for Neural Haircut.
- Reconstruction takes 4-6 hours total versus 72-108 hours for Neural Haircut, roughly tenfold faster.
- Qualitatively reconstructs straight, wavy, and curly hairstyles from monocular video, where Neural Haircut produces over-smoothed or straight strands.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MonoHair instead of Neural Haircut | Monocular-video hair reconstruction · Synthetic dataset [37], 4mm/40° | F-score: 40.3 -> 40.4 | Helped |
| Use MonoHair instead of Neural Haircut | Monocular-video hair reconstruction · Synthetic dataset [37], 2mm/20° | Precision: 52.9 -> 60.8 | Helped |
| Use rendered undirectional strand maps and depth maps with DeepMVSHair* | DeepMVSHair · Synthetic dataset [37], 4mm/40° | F-score: 37.8 -> 38.9 | Helped |
| Add PMVO to reconstruct the exterior geometry (ablation) | MonoHair · Real-world monocular video | hair geometry completeness | Helped |
| Add DeepMVSHair* to infer the internal geometry (ablation) | MonoHair · Real-world monocular video | internal hair structure | Helped |
Surprising findings
- Data-prior-heavy methods such as Neural Haircut over-smooth curly hair, and a prior-free exterior reconstruction recovers those details better.
Limitations
- Connection relationships can be wrong for braids and hairstyles with severe intersections and occlusions.
- Interior structure still depends on the learned data prior from the limited USC-HairSalon database.
Experimental setup
Datasets: USC-HairSalon synthetic hair data [37] H3DS
Benchmarks: synthetic hair data [37] H3DS real-world monocular video captures
Metrics: precision recall F-score reconstruction time
Baselines: Neural Haircut DeepMVSHair DeepMVSHair* NeuS Instant-NGP NeuS-PMVO
Builds on: DeepMVSHair Instant-NGP Neural Haircut PatchMatch MVS for hair (Nam et al.)
Closest papers by meaning
- DiffLocks: Generating 3D Hair from a Single Image using Diffusion Models (CVPR 2025 · similarity 0.91)
- Dr.Hair: Reconstructing Scalp-Connected Hair Strands without Pre-Training via Differentiable Rendering of Line Segments (CVPR 2024 · similarity 0.90)
- Perm: A Parametric Representation for Multi-Style 3D Hair Modeling (ICLR 2025 · similarity 0.89)
- Text-Conditioned Generative Model of 3D Strand-based Human Hairstyles (CVPR 2024 · similarity 0.88)
- DGH: Dynamic Gaussian Hair (NeurIPS 2025 · similarity 0.87)
- GroomLight: Hybrid Inverse Rendering for Relightable Human Hair Appearance Modeling (CVPR 2025 · 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.