KTPFormer: Kinematics and Trajectory Prior Knowledge-Enhanced Transformer for 3D Human Pose Estimation (CVPR2024)

This is the official implementation for "KTPFormer: Kinematics and Trajectory Prior Knowledge-Enhanced Transformer for 3D Human Pose Estimation (CVPR2024)" on PyTorch platform.

KTPFormer: Kinematics and Trajectory Prior Knowledge-Enhanced Transformer for 3D Human Pose Estimation
Jihua Peng, Yanghong Zhou, P. Y. Mok
In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024

Dependencies

Make sure you have the following dependencies installed:

pytorch >= 0.4.0
matplotlib=3.1.0
einops
timm

Dataset

The Human3.6M dataset and HumanEva dataset setting follow the VideoPose3D. Please refer to it to set up the Human3.6M dataset (under ./data directory).

The MPI-INF-3DHP dataset setting follows the P-STMO. Please refer it to set up the MPI-INF-3DHP dataset (also under ./data directory).

Training from scratch

To train our model using the CPN's 2D keypoints as inputs under 243 frames, please run:

python run_ktpformer.py -k cpn_ft_h36m_dbb -f 243 -s 128 -l log/run -c checkpoint

Evaluating

You can download our pre-trained models from Google Drive. Put them in the ./checkpoint directory.

To evaluate our model using the CPN's 2D keypoints as inputs under 243 frames, please run:

python run_ktpformer.py -k cpn_ft_h36m_dbb -c checkpoint --evaluate model_243_CPN_best_epoch.bin -f 243

To evaluate our model using the ground-truth 2D keypoints as inputs under 243 frames, please run:

python run_ktpformer.py -k gt -c checkpoint --evaluate model_243_GT_best_epoch.bin -f 243

Visulization

Please refer to the MHFormer.

Citation

If you find this repo useful, please consider citing our paper:

@inproceedings{peng2024ktpformer,
  title={KTPFormer: Kinematics and Trajectory Prior Knowledge-Enhanced Transformer for 3D Human Pose Estimation},
  author={Peng, Jihua and Zhou, Yanghong and Mok, PY},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={1123--1132},
  year={2024}
}

Acknowledgement

Our code refers to the following repositories.

We thank the authors for releasing their codes.

JihuaPeng / KTPFormer

readme