PoseGraphNet

This is the implementation of the PoseGraphNet model proposed in the paper:

Banik, Soubarna, Alejandro Mendoza GarcÍa, and Alois Knoll. "3D human pose regression using graph convolutional network." 2021 IEEE International Conference on Image Processing (ICIP). IEEE, 2021.

Results on Human3.6M

Model	MPJPE (P1)
PoseGraphNet (Mask R-CNN)	59.5
PoseGraphNet (CPN)	52.8

Dataset Human3.6M

Download Human3.6M from http://vision.imar.ro/human3.6m/ into data_dir. The directory structure should look like this

├── S1
├── S11
├── S5
├── S6
├── S7
├── S8
└── S9
    └── MyPoseFeatures
        ├── D2_Positions
        ├── D3_Positions
        └── D3_Positions_mono

Donwload Mask R-CNN and CPN detections

Refer to VideoPose3D

cd data
wget https://dl.fbaipublicfiles.com/video-pose-3d/data_2d_h36m_cpn_ft_h36m_dbb.npz
wget https://dl.fbaipublicfiles.com/video-pose-3d/data_2d_h36m_detectron_ft_h36m.npz

To train the model

Using CPN Predicted 2D input

python train_posegraphnet_singleloss.py --exp='experiment' --exp_suffix='run1' --run_suffix=1 --exp_desc="description" --data_dir='<DATA_DIR>' --cpn_file='<PATH TO CPN PREDICTED 2D POSE>'

Using ground truth 2D input

set ds_category='gt' in params.json in experiment directory

python train_posegraphnet_singleloss.py --exp='experiment' --exp_suffix='run1' --run_suffix=1 --exp_desc="description" --data_dir='<DATA_DIR>'

To evaluate the model

python train_posegraphnet_singleloss.py --exp='icip_v2' --exp_suffix='run3' --run_suffix='1' --exp_desc='evaluate icip_v2/run3' --test --checkpoint='../models/icip_v2/run3/best.pth.tar' --data_dir='<DATA_DIR>' --cpn_file='<PATH TO CPN PREDICTED 2D POSE>'

Citation

If you use our code, please cite as follow:

@inproceedings{banik20213d, title={3D human pose regression using graph convolutional network}, author={Banik, Soubarna and Garc{'I}a, Alejandro Mendoza and Knoll, Alois}, booktitle={2021 IEEE International Conference on Image Processing (ICIP)}, pages={924--928}, year={2021}, organization={IEEE} }