NormalGAN

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.

[Project Page] [paper]

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).

Changelog

2020.08.11 Release the test code and pretrained models

Requirements

The code and released model were trained on

Ubuntu 16.04 & Python 3.5.2
Pytorch 1.12
trimesh 3.2.36
Python-opencv 3.4.3 (for Python-opencv >= 4.0, please change Line 266 of NormalGAN/src/ops.py to contours, _ = cv2.findContours(mask.astype(np.uint8), cv2.RETR_LIST, cv2.CHAIN_APPROX_NONE))

Optional for src/redner.py:

Pyrender 0.1.32

Recommend for Kinect v2 python implement

pylibfreenect2

Pretrained models

Download the pretrained models in Pretrained models.

Put the pretrained models in NormalGAN/model/pretrained/

Testing on the given images

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.

Testing on your own data

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.

You should first apply the body mask for your RGB-D images before testing them with NormalGAN.
Create NormalGAN/datasets/your_data/color & NormalGAN/datasets/your_data/depth folders, put your own RGB-D data into the folders (use the same filename for your RGB-D image pairs, eg. NormalGAN/datasets/your_data/color/1.png & NormalGAN/datasets/your_data/depth/1.png).
Use NormalGAN/data_utils/createcsv.py to create csv file for your own data.

cd NormalGAN/datasets
python data_utils/createcsv.py your_data_folder_name/ your_csv_file_name.csv
cd ..

Run NormalGAN/test_offline.sh to test your data.

sh test_offline.sh your_csv_file_name.csv your_save_folder_name

Citation

@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}
}

LizhenWangT / NormalGAN