← ArchivePaper2018
HairNet: Single-View Hair Reconstruction Using Convolutional Neural Networks
Yi Zhou, Liwen Hu, Jun Xing, Weikai Chen, Han-Wei Shen, Jianshi Huang, Xin Chen
CNN mapping 2D orientation field input to scalp-parameterized strand features; generates 30K-strand hairstyle in real time with a novel collision loss.
Abstract
We introduce a deep learning-based method to generate full 3D hair geometry from an unconstrained image. Our method can recover local strand details and has real-time performance. State-of-the-art hair modeling techniques rely on large hairstyle collections for nearest neighbor retrieval and then perform ad-hoc refinement. Our deep learning approach, in contrast, is highly efficient in storage and can run 1000 times faster while generating hair with 30K strands. The convolutional neural network takes the 2D orientation field of a hair image as input and generates strand features that are evenly distributed on the parameterized 2D scalp. We introduce a collision loss to synthesize more plausible hairstyles, and the visibility of each strand is also used as a weight term to improve the reconstruction accuracy. The encoder-decoder architecture of our network naturally provides a compact and continuous representation for hairstyles, which allows us to interpolate naturally between hairstyles. We use a large set of rendered synthetic hair models to train our network. Our method scales to real images because an intermediate 2D orientation field, automatically calculated from the real image, factors out the difference between synthetic and real hairs.
How to read this
- Category
- Method: single-view 3D hair reconstruction with a CNN
- Contributions
- Generates full 3D hair geometry (30K strands) from a single unconstrained image in real time, far faster than retrieval-based pipelines
- Uses an encoder-decoder CNN that maps a 2D orientation field to strand features evenly distributed on a parameterized 2D scalp
- Introduces a collision loss for more plausible hairstyles and uses per-strand visibility as a reconstruction weight, with a compact latent space allowing interpolation between hairstyles
- Context
- Departs from database nearest-neighbor retrieval methods such as AutoHair (Chai et al. 2016) by learning a direct image-to-strand mapping, training on rendered synthetic hair and bridging to real images via an intermediate 2D orientation field.Builds on: AutoHair: Fully Automatic Hair Modeling from a Single Image
- Correctness
- Trained on synthetic rendered hair and relying on an automatically computed orientation field, so real-image accuracy hinges on that field and on synthetic-to-real generalization; it recovers plausible strands rather than guaranteed ground-truth geometry, especially for occluded interior hair.
- Clarity
- Accessible for readers with CNN background; a first pass conveys the orientation-field-to-strands idea, and a second pass clarifies the scalp parameterization and the collision/visibility losses.
- How to read it
- Read for the input representation (2D orientation field) and the strand decoder plus collision loss; second pass worthwhile if you work on learned hair or single-view reconstruction.
Related work
- Perm: A Parametric Representation for Multi-Style 3D Hair Modeling 2024 / ICLR
- Cloth and Skin Deformation with a Triangle Mesh Based Convolutional Neural Network 2020 / CGF
- Neural Cloth Simulation 2022 / SIGGRAPH Asia
- A Pixel-Based Framework for Data-Driven Clothing 2020 / SCA
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 →