VIBE: Video Inference for Human Body Pose and Shape Estimation [CVPR-2020]

Check our YouTube videos below for more details.

Paper Video	Qualitative Results

VIBE: Video Inference for Human Body Pose and Shape Estimation,
Muhammed Kocabas, Nikos Athanasiou, Michael J. Black,
IEEE Computer Vision and Pattern Recognition, 2020

Features

Video Inference for Body Pose and Shape Estimation (VIBE) is a video pose and shape estimation method. It predicts the parameters of SMPL body model for each frame of an input video. Pleaser refer to our arXiv report for further details.

This implementation:

has the demo and training code for VIBE implemented purely in PyTorch,
can work on arbitrary videos with multiple people,
supports both CPU and GPU inference (though GPU is way faster),
is fast, up-to 30 FPS on a RTX2080Ti (see this table),
achieves SOTA results on 3DPW and MPI-INF-3DHP datasets,
includes Temporal SMPLify implementation.
includes the training code and detailed instruction on how to train it from scratch.

Getting Started

VIBE has been implemented and tested on Ubuntu 18.04 with python >= 3.7. It supports both GPU and CPU inference. If you don't have a suitable device, try running our Colab demo.

Clone the repo:

git clone https://github.com/mkocabas/VIBE.git

Install the requirements using virtualenv or conda:

# pip
source scripts/install_pip.sh

# conda
source scripts/install_conda.sh

Running the Demo

We have prepared a nice demo code to run VIBE on arbitrary videos. First, you need download the required data(i.e our trained model and SMPL model parameters). To do this you can just run:

source scripts/prepare_data.sh

Then, running the demo is as simple as:

# Run on a local video
python demo.py --vid_file sample_video.mp4 --output_folder output/ --display

# Run on a YouTube video
python demo.py --vid_file https://www.youtube.com/watch?v=wPZP8Bwxplo --output_folder output/ --display

Refer to doc/demo.md for more details about the demo code.

Sample demo output with the --sideview flag:

Google Colab

If you do not have a suitable environment to run this project then you could give Google Colab a try. It allows you to run the project in the cloud, free of charge. You may try our Colab demo using the notebook we have prepared:

Training

Run the commands below to start training:

source scripts/prepare_training_data.sh
python train.py --cfg configs/config.yaml

Note that the training datasets should be downloaded and prepared before running data processing script. Please see doc/train.md for details on how to prepare them.

Evaluation

Here we compare VIBE with recent state-of-the-art methods on 3D pose estimation datasets. Evaluation metric is Procrustes Aligned Mean Per Joint Position Error (PA-MPJPE) in mm.

Models	3DPW ↓	MPI-INF-3DHP ↓	H36M ↓
SPIN	59.2	67.5	41.1
Temporal HMR	76.7	89.8	56.8
VIBE	56.5	63.4	41.5
VIBE + 3DPW	51.9	64.6	41.4

See doc/eval.md to reproduce the results in this table or evaluate a pretrained model.

Citation

@inproceedings{kocabas2019vibe,
  title={VIBE: Video Inference for Human Body Pose and Shape Estimation},
  author={Kocabas, Muhammed and Athanasiou, Nikos and Black, Michael J.},
  booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
  month = {June},
  year = {2020}
}

License

This code is available for non-commercial scientific research purposes as defined in the LICENSE file. By downloading and using this code you agree to the terms in the LICENSE. Third-party datasets and software are subject to their respective licenses.

References

We indicate if a function or script is borrowed externally inside each file. Here are some great resources we benefit:

Pretrained HMR and some functions are borrowed from SPIN.
SMPL models and layer is from SMPL-X model.
Some functions are borrowed from Temporal HMR.
Some functions are borrowed from HMR-pytorch.
Some functions are borrowed from Kornia.
Pose tracker is from STAF.

fred1912/VIBE