This repository contains code for the paper: 3D Human Pose Estimation via Spatial Graph Order Attention and Temporal Body Aware Transformer
- Python 3.8.2
- PyTorch 3.9.16
- CUDA 12.0
We follow the links provided by PoseFormer, GLA-GCN, and P-STMO
- Human3.6m: CPN 2D, Ground_truth 2D, Ground_truth 3D
- HumanEva-I: Ground_truth 2D, Ground_truth 3D
- MPI-INF-3DHP: Train and Test
Please put them in folder ./data to reproduce the results.
We provide the pre-trained models using CPN and GT 2D data, please put them in ./checkpoint directory. To evaluate, pleasae run:
- On ground truth 2D
main_h36m.py -k gt -f 324 --evaluate h36m_gt_best.bin
- On CPN 2D
python main_h36m.py -k cpn_ft_h36m_dbb -f 324 --evaluate h36m_cpn_best.bin
python main_heva.py -k gt -f 5 --evaluate heva1_gt_best.bin
python main_3dhp.py -f 81 -frame-kept 9 -coeff-kept 9 --reload 1 --previous_dir checkpoint/best_mpi-inf-3dhp.pth
- On ground truth 2D
python main_h36m.py -k gt -f 324 -lr 0.0003 -lrd 0.95
- On CPN 2D
python main_h36m.py -k cpn_ft_h36m_dbb -f 324 -lr 0.0003 -lrd 0.95
python main_heva.py -k gt -f 5 -lr 0.001 -lrd 0.95
python main_3dhp.py -f 81 -frame-kept 9 -coeff-kept 9 -b 512 --train 1 --lr 0.0007 -lrd 0.97 -c CKPT_NAME --gpu 1