Gradient Multi-Foci Networks for 3D skeleton-based Human Motion Prediction

This is the code for the paper

Junyu Shi, Jianqi Zhong, Zhiquan He, Wenming Cao. "Gradient Multi-Foci Networks for 3D skeleton-based Human Motion Prediction."

Dependencies

• cuda 12.1
• Python 3.10.0
• Pytorch =2.1.0

Get the data

Human3.6m in exponential map can be downloaded from here.

Directory structure:

H3.6m
|-- S1
|-- S5
|-- S6
|-- ...
`-- S11

Training

Training on Human3.6M

python main_h36m_3d.py --kernel_size 10 --dct_n 20 --input_n 50 --output_n 10 --skip_rate 1 --batch_size 32 --test_batch_size 256 --in_features 66 --lr 0.001 --dev cuda:0 --data_dir [PATH TO DATA]

Evaluation

Testing on Human3.6M

python main_h36m_3d_eval.py --is_eval --kernel_size 10 --dct_n 20 --input_n 50 --output_n 10 --skip_rate 1 --test_batch_size 256 --in_features 66 --ckpt [PATH TO CKPT] --dev cuda:0 --data_dir [PATH TO DATA]

To do

• Release the training code on Human3.6M dataset
• Release the training code on amass dataset
• Release the training code on CMU dataset

Acknowledgments

The overall code framework (dataloading, training, testing etc.) is based on HisRepItself.

Licence

MIT