GroomCap: Prior-Free Multi-View Hair Capture

• GroomCap is a prior-free multi-view hair capture method that reconstructs strand-level hair geometry using neural implicit representations.
• It integrates volumetric orientation rendering and Gaussian-based optimization to preserve intricate hair details and reduce smoothing artifacts.
• The approach produces explicit 3D hair models suitable for VR, digital avatars, VFX, and interactive simulations.

GroomCap is a prior-free, multi-view hair capture methodology that reconstructs high-fidelity, strand-level hair geometry from images, without reliance on external data priors. Designed to overcome the intrinsic challenges and smoothing artifacts of conventional algorithms, GroomCap advances the state of 3D hair modeling by integrating neural implicit representations for volumetric hair, a volumetric orientation rendering and supervision framework, and an explicit Gaussian-based strand optimization strategy. This architecture facilitates the reconstruction of highly detailed, explicit hair models adaptable to a wide array of digital and computational graphics applications (Zhou et al., 1 Sep 2024).

1. Conceptual Foundations and Motivation

Traditional multi-view hair reconstruction methods employ explicit surfaces or prior-driven models, often leading to loss of strand-level structure, excessive smoothing, and poor generalization to diverse hairstyles or subject-specific features. They may fail to capture overlapping or intricately clustered arrangements (e.g., flyaway strands, hairlines).

GroomCap addresses these limitations by eliminating dependence on synthetic or external datasets. Instead, it represents the volumetric hair field as an implicit neural function trained directly on multi-view images. This approach is motivated by the need for a flexible system capable of preserving intricate, subject-specific geometry and appearance features, while retaining theoretical generality for deployment across virtual reality, gaming, and photorealistic avatar pipelines.

2. Neural Implicit Hair Volume Representation

GroomCap employs a neural implicit volume—specifically, an MLP that encodes the hair field by mapping a 3D spatial point and optional view direction to several outputs:

• Volume density ($\sigma$)
• Hair occupancy ($\rho_h$) and body occupancy ($\rho_b$)
• 3D hair orientation in undirectional polar angles $(\theta, \phi)$

Adjunct to geometry, a view-dependent appearance MLP predicts radiance, analogous to a volumetric radiance field (e.g., NeRF).

A critical innovation targets the deficiency of conventional volumetric approaches, which, when aggregating orientations along viewing rays, inherently average overlapping strands, causing orientation blending artifacts. To prevent this, GroomCap expands each predicted hair orientation at a spatial location into a local probability density function on ($\theta$, $\phi$)-space via a kernel:

$$h(\theta, \phi) = \frac{1}{C} \cdot h'(\theta, \phi)\,,$$

where

$$h'(\theta, \phi) = \frac{1}{\beta ( \lVert \theta - \theta_0 \rVert^2 + \lVert \phi - \phi_0 \rVert^2 ) + \delta}$$

and $$C = \iint_0^\pi h'(\theta, \phi) \, d\theta\, d\phi$$. Multiple shifts are summed to enforce the required periodicity for undirectional hair. The orientation distributions are then accumulated along a camera ray $r$ to yield

$$g_r(\theta, \phi) = \int_{t_n}^{t_f} T(t) \sigma(r(t)) h_{r(t)}(\theta, \phi) dt$$

with $$T(t) = \exp \left( - \int_{t_n}^{t} \sigma(r(\alpha)) d\alpha \right)$$. This formulation preserves per-strand information even in densely overlapping regions.

3. Orientation Distribution Supervision and Loss

Supervision is imposed not on single best-fit angles, but on projected 2D orientation distributions derived from the images. Each source view is filtered with a Gabor filterbank to estimate per-pixel local orientation distributions $f(\eta)$. The predicted distribution $\bar{f}(\eta)$ is computed from the rendered orientation volume.

The loss penalizes squared $L^2$ distance between predicted and observed orientation distributions:

$$\mathcal{L}_{\mathrm{ori}} = \int_0^\pi \| f(\eta) - \bar{f}(\eta) \|^2 d\eta$$

This volumetric orientation rendering algorithm prevents mode collapse and loss of structure that would occur if orientations were simply averaged.

4. Gaussian-Based Hair Strand Optimization

Initial hair strands are traced from the implicit volume using a forward Newton solver. These polylines are refined using a chained Gaussian representation: every segment (between vertices $v_i$ and $v_{i+1}$) is represented by a Gaussian with its covariance $C$ computed as

$C = E D D^\top E^\top$

where $E$ contains principal axes (primary axis aligned with the segment, secondaries transverse), and $D$ is diagonal with elements $\tau_l = \| v_{i+1} - v_i \| / 2$ (axial) and $\tau_d = d/2$ (radial, with $d$ the strand diameter).

Differentiable rasterization is performed via Gaussian splatting, enabling photometric losses directly between synthetic and observed images. Hair parameter explosion is averted by representing each strand’s appearance and geometry with a small set of anchor parameters, regularized through a strand-VAE that imposes a low-dimensional latent prior. During optimization, adaptive control—including strand splitting (weighted by a split score computed over opacity-weighted length contributions) and pruning—ensures variable strand density matching observed structure.

5. Experimental Results and Comparative Analysis

GroomCap reconstructs explicit scalp-rooted polylines, typically yielding $\sim$150,000 strands per subject (each with 100 vertices). Qualitative results reported for both studio-captured and in-the-wild subjects (e.g., on Neural Haircut data) show that GroomCap exceeds prior works (MonoHair, Neural Haircut) in maintaining multi-layered, overlapping structures, nuanced hairlines, and subject-specific topologies.

Ablation studies indicate the necessity of full orientation distribution supervision and Gaussian-based optimization for effective strand separation and geometric realism. Quantitative accuracy scores (e.g., mean error metrics) are not explicitly reported, but visual evidence and comparison suggest substantial increases in fidelity over previous algorithms.

6. Application Domains

Because output is an explicit geometric hair model, GroomCap’s representation supports:

• Physically based rendering: Enables detailed relighting and appearance editing.
• Simulation: Scalped-rooted strands with preserved topology facilitate integration with hair simulation engines.
• Interactive editing: Artists and users can manipulate individual strands for grooming or virtual haircut scenarios.

These attributes position GroomCap for broad utility in digital avatar creation, feature film VFX, VR contexts, and high-end gaming.

7. Limitations and Future Directions

The authors note that, while GroomCap significantly enhances prior-free hair modeling, it may be challenged by scenes with extremely dark, curly, or otherwise visually ambiguous hair, given the lack of external priors. Combining data-driven priors with the implicit/explicit approach of GroomCap represents a direction for further research, with prospective benefit for robustness in severe cases and improved segment detection and density estimation. This suggests that hybrid methods may emerge as a future standard in high-fidelity hair capture pipelines.

In summary, GroomCap introduces a mathematically grounded, data-efficient, and explicit approach for high-fidelity multi-view hair capture, constructing structured, photorealistic 3D hair suitable for flexible, physically accurate applications in digital graphics and beyond (Zhou et al., 1 Sep 2024).