"Learning Semantic Latent Directions for Accurate and Controllable Human Motion Prediction" (ECCV 2024)

---

This repo contains the official implementation of the paper:

Learning Semantic Latent Directions for Accurate and Controllable Human Motion Prediction

ECCV 2024 [arxiv]

Dependencies

• Python >= 3.8
• PyTorch >= 1.9
• Tensorboard
• matplotlib
• tqdm
• argparse

Get the data

We adapt the data preprocessing from GSPS.

• We follow the data preprocessing steps (DATASETS.md) inside the VideoPose3D repo.
• Given the processed dataset, we further compute the multi-modal future for each motion sequence. All data needed can be downloaded from Google Drive and place all the dataset in data folder inside the root of this repo.

Get the pretrain models

• All pretrain models can be downloaded from Google Drive and place all the pretrain models in results folder inside the root of this repo.

Train

We have used the following commands for training the network on Human3.6M or HumanEva-I with skeleton representation:

python train_nf.py --cfg [h36m/humaneva] --gpu_index 0
python main.py --cfg [h36m/humaneva] --gpu_index 0

Test

To test on the pretrained model, we have used the following commands:

python main.py --cfg [h36m/humaneva] --mode test --iter 500 --gpu_index 0

Visualization

For visualizing from a pretrained model, we have used the following commands:

python main.py --cfg [h36m/humaneva] --mode viz --iter 500 --gpu_index 0

Acknowledgments

This code is based on the implementations of STARS.

Citation

If you find this work useful in your research, please cite:

@article{xu2024learning,
  title={Learning Semantic Latent Directions for Accurate and Controllable Human Motion Prediction},
  author={Xu, Guowei and Tao, Jiale and Li, Wen and Duan, Lixin},
  journal={arXiv preprint arXiv:2407.11494},
  year={2024}
}

License

This repo is distributed under an MIT LICENSE