[CVPR2023] Generating Holistic 3D Human Motion from Speech
[Project Page] [Arxiv] [Colab]
This repository provides the official implementation of SHOW(Synchronous HOlistic body in the Wild). Given rgb images or videos only, SHOW can reconstruct holistic whole body mesh results. Please refer to the arXiv paper for more details.
What you can use:
- easy and efficient: adapts SMPLify-X to the videos of talking persons with several good practices.
- state-of-art: Compared to other methods, ours produces more accurate and stable results with details.
Take a quick tour on colab: [Colab].
another [colab] version from @SlimeVRX
To install SHOW, please execute pip install git+https://github.com/yhw-yhw/SHOW.git or
git clone https://github.com/yhw-yhw/SHOW.git
cd SHOW && pip install -v -e .-
[environment] Using virtual environment by runing
conda create -n env_SHOW python=3.9 eval "$(conda shell.bash hook)" conda activate env_SHOW
install pytorch using
pip3 install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu116orconda install pytorch torchvision torchaudio pytorch-cuda=11.6 -c pytorch -c nvidia
You can run
cd SHOW cd modules/MICA && pip install -r requirements.txt cd ../PIXIE && pip install -r requirements.txt cd ../PyMAF && pip install -r requirements.txt cd ../DECA && pip install -r requirements.txt cd ../.. && pip install -r requirements.txt
Note that Pytorch3D may require manuall installation (see instructions here).
-
[models] download packed model files, and put it in the same level directory as SHOW
wget "https://www.dropbox.com/scl/fi/gwvp5c3yijkjc726bidxx/models.zip?rlkey=2p4m788qpi04oye3kur2pxszx&st=dchhjclv&dl=0" -O models.zip wget "https://www.dropbox.com/scl/fi/vcav90wzwqxmg56n42gr1/data.zip?rlkey=5oetna909azec027v42ogx42q&st=e5mnsldy&dl=0" -O data.zip unzip data.zip 7za x models.zip
-
[OpenPose]: follow the code from [OpenPose Colab notebook](https://colab.research.google.com/github/tugstugi/dl-colab-notebooks/blob/master/notebooks/OpenPose.ipynb, and change OpenPose bin path
openpose_root_pathandopenpose_bin_pathinconfigs\configs\machine_cfg.py. -
[MMPose]: Make sure to install mmcv-full, and set env variable
mmpose_rootpip install openmim mim install mmcv-full git clone https://github.com/open-mmlab/mmdetection cd /content/mmdetection && python setup.py install git clone https://github.com/open-mmlab/mmpose export mmpose_root = $mmpose_root$
-
models for
inisghtface:use the following command as reference
mkdir -p ~/.insightface/models cd ~/.insightface/models wget https://keeper.mpdl.mpg.de/f/2d58b7fed5a74cb5be83/?dl=1 -O antelopev2.zip wget https://keeper.mpdl.mpg.de/f/8faabd353cfc457fa5c5/?dl=1 -O buffalo_l.zip mkdir -p antelopev2 && cd antelopev2 && unzip -o ../antelopev2.zip mkdir -p buffalo_l && cd buffalo_l && unzip -o ../buffalo_l.zip
-
[others]
pip uninstall -y xtcocotools && pip install xtcocotools --no-binary xtcocotools
Run SHOW on demo video
python main.py --speaker_name -1 --all_top_dir ./test/demo_video/half.mp4It takes 15-20 minutes for 5s 30 FPS video on Colab Tesla T4.
The final results are stored in ./test/demo_video/ours_exp. All the smplx variables can be found in ./test/demo_video/ours_exp/final_all.pkl, and the visualization can be viewed in ./test/demo_video/ours_exp/final_vis.mp4.
Download all videos from youtube, please refer to (https://github.com/amirbar/speech2gesture), or using the following script: download_youtube.py, remember to install yt-dlp. After downloading all videos, you can using SHOW_intervals_subject4.csv for video interval cropping. oOr download it from https://www.dropbox.com/sh/k3k5iq0hg0y1mjm/AAAar3ObkFPkP4c87sDMggrta?dl=0
After running SHOW, we will get processed data in a pkl file. Here we can visualize pkl from our provided dataset.
python render_pkl_release.py \
--pkl_file_path test/demo_pkl/all.pkl \
--out_images_path test/demo_pkl/ours_images \
--output_video_path test/demo_pkl/ours.mp4 \
--smplx_model_path ../models/smplx/SMPLX_NEUTRAL_2020_org.npzThe data reconstructed by SHOW is released, you can download it
-
speaker=oliver/chemistry/conan/seth
-
The maximum length of video clip is 10s with 30 fps
-
Format of files in the compressed package:
-
{speaker}_wav_tar.tar.gz:- The path format of each file is:
speaker/video_fn/seq_fn.wav - Audio obtained from the original video at 22k sampling rate
- The path format of each file is:
-
{speaker}_pkl_tar.tar.gz:- The path format of each file is:
speaker/video_fn/seq_fn.pkl - Data contained in the pkl file:
width，height: the video width and height center: the center point of the video batch_size: the sequence length camera_transl: the displacement of the camera focal_length: the pixel focal length of a camera body_pose_axis: (bs, 21, 3) expression: (bs, 100) jaw_pose: (bs,3) betas: (300) global_orient: (bs,3) transl: (bs,3) left_hand_pose: (bs,12) right_hand_pose: (bs,12) leye_pose: (bs,3) reye_pose: (bs,3) pose_embedding: (bs,32)
- Set the config of smplx model as follows:
smplx_cfg=dict( model_path='path_to_smplx_model' model_type= 'smplx', gender= 'neutral', use_face_contour= True, use_pca= True, flat_hand_mean= False, use_hands= True, use_face= True, num_pca_comps= 12, num_betas= 300, num_expression_coeffs= 100, )
- The path format of each file is:
-
-
In practice, global orient and transl parameters should be fixed as the first frame and the lower part of the body pose should be fixed as sitting or standing position: code
usage
python cvt_exp_dim_tool.py \
--target-exp-dim 50 \
--pkl-path ./rich.npy \
--model-path ../models/smplx/SMPLX_MALE_shape2019_exp2020.npz
If you use this project in your research please cite this paper:
@inproceedings{yi2022generating,
title={Generating Holistic 3D Human Motion from Speech},
author={Yi, Hongwei and Liang, Hualin and Liu, Yifei and Cao, Qiong and Wen, Yandong and Bolkart, Timo and Tao, Dacheng and Black, Michael J},
booktitle={CVPR},
year={2023}
}
-
If following error is encountered
RuntimeError: Subtraction, the `-` operator, with a bool tensor is not supported. If you are trying to invert a mask, use the `~` or `logical_not()` operator instead.Open torchgeometry/core/conversions.py file and replace line 304-306 with
mask_c1 = mask_d2 * (~ mask_d0_d1) mask_c2 = (~ mask_d2) * mask_d0_nd1 mask_c3 = (~ mask_d2) * (~ mask_d0_nd1)
This code and model are available for non-commercial scientific research purposes as defined in the LICENSE file. By downloading and using the code and model you agree to the terms in the LICENSE.
For functions or scripts that are based on external sources, we acknowledge the origin individually in each file. We specifically thanks WOJCIECH ZIELONKA and Justus Thies for sharing their face tracking codebase.
Here are some great resourc 716D es we benefit:
- SMPLify-X
- DECA for face expression initialization
- PIXIE and PyMAF-X for SMPL-X parameters initialization
- DeepLab for person segmentation
- MICA and [https://github.com/HavenFeng/photometric_optimization] for face tracking
- MICA_Tracker
- Pytorch3D for rendering
- FAN for landmark detection
- arcface-pytorch
For questions, please contact talkshow@tue.mpg.de or hongwei.yi@tuebingen.mpg.de or fthualinliang@mail.scut.edu.cn.
For commercial licensing, please contact ps-licensing@tue.mpg.de