NormalGAN: Learning Detailed 3D Human from a Single RGB-D Image (ECCV 2020)
Lizhen Wang, Xiaochen Zhao, Tao Yu, Songtao Wang and Yebin Liu
We propose NormalGAN, a fast adversarial learning-based method to reconstruct the complete and detailed 3D human from a single RGB-D image.
Note: As we can not release our dataset, we do not release the training code. Now you can try NormalGAN on 3D dataset like THUman2.0. If you are interested with our training code, please fell free to send an e-mail to Lizhen Wang (wlz18@mails.tsinghua.edu.cn).
2020.08.11 Release the test code and pretrained models
The code and released model were trained on
contours, _ = cv2.findContours(mask.astype(np.uint8), cv2.RETR_LIST, cv2.CHAIN_APPROX_NONE)
)Optional for src/redner.py:
Recommend for Kinect v2 python implement
Download the pretrained models in Pretrained models.
Put the pretrained models in NormalGAN/model/pretrained/
Generate the csv file for the demo images in NormalGAN/datasets/testdata.
cd NormalGAN/datasets
python data_utils/createcsv.py testdata/ test.csv
cd ..
Run the NormalGAN/test_offline.sh file (which occupies about 3.5-GB GPU memory).
sh test_offline.sh test.csv testdata
Results are shown in NormalGAN/results/testdata/ply. You can also use Poisson Reconstruction for better performance of the edge area.
Please note that, NormalGAN simulate noise for Kinect v2 (or similar ToF depth cameras), the image resolution should be (512,424). Please change the camera intrinsics and image resolution in NormalGAN/src/test_offline.py.
cd NormalGAN/datasets
python data_utils/createcsv.py your_data_folder_name/ your_csv_file_name.csv
cd ..
sh test_offline.sh your_csv_file_name.csv your_save_folder_name
@inproceedings{wang2020normalgan,
title={NormalGAN: Learning Detailed 3D Human from a Single RGB-D Image},
author={Wang, Lizhen and Zhao, Xiaochen and Yu, Tao and Wang, Songtao and Liu, Yebin},
booktitle={ECCV},
year={2020}
}