PIFuHD: Multi-Level Pixel-Aligned Implicit Function for High-Resolution 3D Human Digitization (CVPR 2020)

News:

• [2020/06/15] Demo with Google Colab (incl. visualization) is available! Please check out #pifuhd on Twitter for many results tested by users!

This repository contains a pytorch implementation of "Multi-Level Pixel-Aligned Implicit Function for High-Resolution 3D Human Digitization".

This codebase provides:

• test code
• visualization code

See our blog post to learn more about our work at CVPR2020!

Demo on Google Colab

In case you don't have an environment with GPUs to run PIFuHD, There is a Google Colab demo. You can also upload your own images and reconstruct 3D geometry together with visualization. Try the Colab demo using the following notebook:

Requirements

• Python 3
• PyTorch tested on 1.4.0, 1.5.0
• json
• PIL
• skimage
• tqdm
• cv2

For visualization

• trimesh with pyembree
• PyOpenGL
• freeglut (use sudo apt-get install freeglut3-dev for ubuntu users)
• ffmpeg

Note: At least 8GB GPU memory is recommended to run PIFuHD model.

Run the following code to install all pip packages:

pip install -r requirements.txt 

Download Pre-trained model

Run the following script to download the pretrained model. The checkpoint is saved under ./checkpoints/.

sh ./scripts/download_trained_model.sh

A Quick Testing

To process images under ./sample_images, run the following code:

sh ./scripts/demo.sh

The resulting obj files and rendering will be saved in ./results. You may use meshlab (http://www.meshlab.net/) to visualize the 3D mesh output (obj file).

Testing

1. run the following script to get joints for each image for testing (joints are used for image cropping only.). Make sure you correctly set the location of OpenPose binary. Alternatively colab demo provides more light-weight cropping rectange estimation without requiring openpose.

python apps/batch_openpose.py -d {openpose_root_path} -i {path_of_images} -o {path_of_images}

2. run the following script to run reconstruction code. Make sure to set --input_path to path_of_images, --out_path to where you want to dump out results, and --ckpt_path to the checkpoint. Note that unlike PIFu, PIFuHD doesn't require segmentation mask as input. But if you observe severe artifacts, you may try removing background with off-the-shelf tools such as removebg. If you have {image_name}_rect.txt instead of {image_name}_keypoints.json, add --use_rect flag. For reference, you can take a look at colab demo.

python -m apps.simple_test

3. optionally, you can also remove artifacts by keeping only the biggest connected component from the mesh reconstruction with the following script. (Warning: the script will overwrite the original obj files.)

python apps/clean_mesh.py -f {path_of_objs}

Visualization

To render results with turn-table, run the following code. The rendered animation (.mp4) will be stored under {path_of_objs}.

python -m apps.render_turntable -f {path_of_objs} -ww {rendering_width} -hh {rendering_height} 
# add -g for geometry rendering. default is normal visualization.

Citation

@inproceedings{saito2020pifuhd,
  title={PIFuHD: Multi-Level Pixel-Aligned Implicit Function for High-Resolution 3D Human Digitization},
  author={Saito, Shunsuke and Simon, Tomas and Saragih, Jason and Joo, Hanbyul},
  booktitle={CVPR},
  year={2020}
}

License

See the LICENSE file.