Home

Awesome

Probabilistic-Monocular-3D-Human-Pose-Estimation-with-Normalizing-Flows

This is the official implementation of the ICCV 2021 Paper "Probabilistic Monocular 3D Human Pose Estimation with Normalizing Flows" by Tom Wehrbein, Marco Rudolph, Bodo Rosenhahn and Bastian Wandt.

Installation instructions

We recommend creating a clean conda environment. You can do this as follows:

conda env create -f environment.yml

After the installation is complete, you can activate the conda environment by running:

conda activate ProHPE

Finally, FrEIA needs to be installed:

pip install git+https://github.com/VLL-HD/FrEIA.git@ad2f5a261a2fc0002fb4c9adeff7a62b0e9dd4e1

Data

Download the used Human3.6M detections and Gaussian fits from Google Drive. Afterwards, extract the zip file to the data/ directory.

Due to licensing, it is not possible for us to provide any data from Human3.6M. Therefore, the 3D pose data needs to be downloaded from the official dataset website (account required). For each subject, download the file 'D3_Positions_mono' and extract it to the data/ directory. Afterwards, run

python create_data.py

to automatically merge the detections and Gaussian fits with the 3D ground truth data.

For information about the dataset structure, see data/data_info.txt.

Run evaluation code

Evaluate the pretrained model on the whole testset of Human3.6M:

python eval_action_wise.py --exp original_model

Evaluate on the hard subset of Human3.6M containing highly ambiguous examples:

python eval_hard_subset.py --exp original_model

Evaluation results are saved to the results/original_model/ directory.

Run training code

Training can be started with:

python train.py --exp experiment_name

Citation

Please cite the paper in your publications if it helps your research:

@inproceedings {WehRud2021,
  author = {Tom Wehrbein and Marco Rudolph and Bodo Rosenhahn and Bastian Wandt},
  title = {Probabilistic Monocular 3D Human Pose Estimation with Normalizing Flows},
  booktitle = {International Conference on Computer Vision (ICCV)},
  year = {2021},
  month = oct
}

Link to the paper: