Hair20K
INTRODUCTION
We introduce a 3D strand-based hair dataset containing 21,054 hairstyles. The dataset was collected for training [PERM], a parametric 3D hair generator, which is applied in [Digital Salon], a real-time interactive hair simulator. The data were created by augmenting the 343 hair models from [USC-HairSalon], using the hair blending methods introduced in [PERM], [HairNet] and [Hu et al.].
In the database,
- Hair styles are organized in 10 catergories;
- We align all the hair models with two standard head models;
- We store each hair model in .data file format, the same as [USC-HairSalon].
- Each hair model constains ~10,000 hair strands. We do not store strands with zero lengths.
- Each strand is sampled as polyline with uniform segment length.
- For getting hair models with uniformly sampled roots on the scalp and uniform segment numbers on the strand, please refer to the code of PERM.
- The database links below contains 10,527 hair models. For mirroring the hair styles to get the complete database, please use the script in PERM.
MAIN PROCESS
We classify the 343 hair models in USC-HairSalon into 10 catergories. Visualizations of the 10 catergories can be found in HairNet_DataSetGeneration/HairClasses. Within each class, we blend each pair of two different hair styles. For blending the hairstyles, we first compute five guidestrands for each hairstyle, with each guidestrand representing a partial area of the hair. Then we randomly pick five guidestrands from the ten guidestrands and form a new hair by blending the five local regions. Blending is done in the volumetric space. After blending, we clean the strands and place the hair on the AVATAR NEO head based on the warping between two heads.
TERMS OF USE
NON-COMMERCIAL USE: The license granted is for internal, non-commercial research, evaluation or testing purposes only. Any use of the DATA or its contents to manufacture or sell products or technologies (or portions thereof) either directly or indirectly for any direct or indirect for-profit purposes is strictly prohibited.
NO WARRANTY: The data are provided "as is" and any express or implied warranties are disclaimed.
NO DISTRIBUTION: You are not allowed to sublicense or distribute the DATA in whole or in part to any third party.
CITATION: You agree to reference the publication "Perm: A Parametric Representation for Multi-Style 3D Hair Modeling" [1], "HairNet: Single-View Hair Reconstruction Using Convolutional Neural Networks." [2] and "Single-View Hair Modeling Using A Hairstyle Database." [3] in any publications reporting use of it or any manual or document. A copy of all reports and papers that are for public or general release that use the DATA must be forwarded immediately upon release or publication to the DATA Principal Investigator.
LINKS
[Data] aligned with the AVATAR NEO head (used for training PERM). TODO: move data to S3.
[Data] aligned with the USC-HairSalon head. TODO: move data to S3.
[AVATAR NEO] head. TODO: ask Hao to release the head model.
[Code] for loading data in python.
[Code] for loading data in C++. Ask Liwen to restore the c++ loading code.
REFERENCES
[1] Perm: A Parametric Representation for Multi-Style 3D Hair Modeling.
2024. Chengan He, Xin Sun, Zhixin Shu, Fujun Luan,Soren Pirk, Jorge Alejandro Amador Herrera, Dominik L. Michels, Tuanfeng Y. Wang, Meng Zhang, Holly Rushmeier and Yi Zhou.
arXiv preprint arXiv::2407.19451
[ Project page ]
[2] HairNet: Single-View Hair Reconstruction Using Convolutional Neural Networks.
2018. Yi Zhou, Liwen Hu, Jun Xing, Weikai Chen, Han-Wei Kung, Xin Tong and Hao Li.
Proceedings of the European Conference on Computer Vision (ECCV) 2018 [paper] [video]
[3] Single-View Hair Modeling Using A Hairstyle Database.
2015. Liwen Hu, Chongyang Ma, Linjie Luo, Hao Li.
ACM Transactions on Graphics (Proceedings of SIGGRAPH 2015), Vol 34, Issue 4, 125:1--125:9.
[ Project page ]