← ArchivePaper2019
Dynamic Hair Modeling from Monocular Videos Using Deep Neural Networks
Two-network framework (HairSpatNet + HairTempNet) inferring 3D occupancy and orientation fields from monocular video to model moving hairstyles dynamically.
Abstract
We introduce a deep learning based framework for modeling dynamic hairs from monocular videos, which could be captured by a commodity video camera or downloaded from Internet. The framework mainly consists of two neural networks, i.e., HairSpatNet for inferring 3D spatial features of hair geometry from 2D image features, and HairTempNet for extracting temporal features of hair motions from video frames. The spatial features are represented as 3D occupancy fields depicting the hair volume shapes and 3D orientation fields indicating the hair growing directions. The temporal features are represented as bidirectional 3D warping fields, describing the forward and backward motions of hair strands cross adjacent frames. Both HairSpatNet and HairTempNet are trained with synthetic hair data. The spatial and temporal features predicted by the networks are subsequently used for growing hair strands with both spatial and temporal consistency. Experiments demonstrate that our method is capable of constructing plausible dynamic hair models that closely resemble the input video, and compares favorably to previous single-view techniques.
How to read this
- Category
- Method: deep learning for dynamic hair capture
- Contributions
- Introduces a two-network framework for modeling dynamic hair from monocular video
- HairSpatNet infers 3D spatial features (occupancy fields for volume shape and orientation fields for growing direction) from 2D image features, and HairTempNet extracts temporal features as bidirectional 3D warping fields
- Grows hair strands with spatial and temporal consistency, comparing favorably to prior single-view techniques
- Context
- Extends single-view hair reconstruction such as HairNet (Zhou et al., 2018) from static stills to temporally consistent dynamic hair from video.Builds on: HairNet: Single-View Hair Reconstruction Using Convolutional Neural Networks
- Correctness
- Both networks are trained on synthetic hair data, so real-video performance depends on the synthetic-to-real gap; results are plausible reconstructions that resemble the input rather than measured ground-truth strand geometry, and monocular ambiguity persists.
- Clarity
- Accessible framing of a two-network split; a first pass conveys the spatial-plus-temporal decomposition, with a second pass for the field representations and warping.
- How to read it
- Read for the spatial/temporal network split and the occupancy-plus-orientation-plus-warping representation; do a second pass on training data and the field formulations to gauge how it transfers to real footage.
Built upon by
Nothing yet.
Related work
- Neural Haircut: Prior-Guided Strand-Based Hair Reconstruction 2023 / ICCV
- BlendSim: Simulation on Parametric Blendshapes using Spacetime Projective Dynamics 2025 / CGF
- Learning Motion Manifolds with Convolutional Autoencoders 2015 / SIGGRAPH Asia
- 3DGS-Avatar: Animatable Avatars via Deformable 3D Gaussian Splatting 2024 / CVPR
Keywords
This page summarises the entry and links to its original source. The archive never hosts or redistributes the publication itself.Show it in the full archive list →