Iterative Graph Filtering Network for 3D Human Pose Estimation (Journal of Visual Communication and Image Representation, 2023)
This repository contains the official PyTorch implementation of the Iterative Graph Filtering Network for 3D Human Pose Estimation authored by Zaedul Islam and A. Ben Hamza. If you discover our code to be valuable for your research, kindly consider including the following citation:
@article{islam2023iterative,
title={Iterative graph filtering network for 3D human pose estimation},
author={Islam, Zaedul and Hamza, A Ben},
journal={Journal of Visual Communication and Image Representation},
volume={95},
pages={103908},
year={2023},
publisher={Elsevier}
}
Results under Protocol #1 (mean per-joint position error) and Protocol #2 (mean per-joint position error after rigid alignment).
| Method | MPJPE (P1) | PA-MPJPE (P2) |
|---|---|---|
| SemGCN | 57.6mm | - |
| High-order GCN | 55.6mm | 43.7mm |
| HOIF-Net | 54.8mm | 42.9mm |
| Weight Unsharing | 52.4mm | 41.2mm |
| MM-GCN | 51.7mm | 40.3mm |
| Modulated GCN | 49.4mm | 39.1mm |
| Ours | 47.1mm | 38.7mm |
This repository is built upon Python v3.8 and Pytorch v1.8.0 on Ubuntu 20.04.4 LTS. All experiments are conducted on a single NVIDIA RTX 3070 GPU with 8GB of memory.
Please make sure you have the following dependencies installed:
- PyTorch >= 1.8.0
- NumPy
- Matplotlib
Our model is evaluated on Human3.6M and MPI-INF-3DHP datasets. Please put the datasets in ./dataset directory.
2D detections for Human3.6M dataset are provided by VideoPose3D Pavllo et al.
We set up the MPI-INF-3DHP dataset in the same way as PoseAug. Please refer to DATASETS.md to prepare the dataset file.
To initiate the training of our model, utilizing the identified 2D keypoints (HR-Net) along with pose refinement, please execute the following command:
python main_graph.py --pro_train 1 --beta 0.2 --k hrn --batchSize 512 --hid_dim 384 --save_model 1 --save_dir './checkpoint/train_result/' --post_refine --save_out_type post --show_protocol2To initiate the training of our model, utilizing the ground truth 2D keypoints excluding pose refinement and non-local layer, please execute the following command:
python main_graph.py --pro_train 1 --beta 0.2 --k gt --batchSize 512 --hid_dim 384 --save_model 1 --save_dir './checkpoint/train_result/' --show_protocol2To evaluate our model using the detected 2D keypoints (HR-Net) with pose refinement, please run the following command:
python main_graph.py -k hrn --beta 0.2 --batchSize 512 --hid_dim 384 --previous_dir './checkpoint/train_result/' --save_out_type post --save_dir './checkpoint/test_result/' --gsnet_gcn_reload 1 --module_gsnet_model [model_gsnet].pth --post_refine --post_refine_reload 1 --post_refine_model `[model_post_refine]`.pth --show_protocol2 --nepoch 2To evaluate our model using the ground truth 2D keypoints without incorporating pose refinement and and non-local layer, please run the following command:
python main_graph.py -k gt --beta 0.2 --batchSize 512 --hid_dim 384 --previous_dir './checkpoint/train_result/' --save_dir './checkpoint/test_result/' --save_out_type xyz --gsnet_gcn_reload 1 --module_gsnet_model `[model_gsnet]`.pth --show_protocol2 --nepoch 2The pre-trained models can be downloaded from Google Drive. Put them in the ./checkpoint/train_result directory.
To evaluate our pre-trained model using the ground truth 2D keypoints, please run:
python main_graph.py -k gt --beta 0.2 --batchSize 512 --hid_dim 384 --previous_dir './checkpoint/train_result/' --save_dir './checkpoint/test_result/' --save_out_type xyz --gsnet_gcn_reload 1 --module_gsnet_model model_gsnet_gcn_7_eva_xyz_3649.pth --show_protocol2 --nepoch 2Our code makes references to the following repositories.
We thank the authors for sharing their code and kindly request that you also acknowledge their contributions by citing their work.