deterministic

Deterministic Human Motion Forecasting on Human3.6M

---

Dependencies

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

Get the data

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

Directory structure:

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

Put the all downloaded datasets in ./data directory.

Train

The arguments for running the code are defined in parser.py. We have used the following commands for training the network on Human3.6M with skeleton representation:

python main.py --input_n [number of historical sequence frames] --output_n [maximum number of predicted frames] --skip_rate [sampling rate] --n_pre [number of dct coefficients] --data_dir ./data --version [long / short]

It's able to predict the future motion considering the global translation

python main.py --input_n [number of historical sequence frames] --output_n [maximum number of predicted frames] --skip_rate [sampling rate] --n_pre [number of dct coefficients] --global_translation --data_dir ./data --version [long / short]

We provide the pretrained model with 10 historical sequence frames and 25 future predicted frames following the literature.

Test

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

python main.py --input_n [number of historical sequence frames] --output_n [maximum number of predicted frames] --test_output_n [index of the test frame] --skip_rate [sampling rate] --n_pre [number of dct coefficients] --mode test --model_path ./checkpoints/CKPT_3D_H36M --data_dir ./data --version [long / short]

With global translation, we have the following commands:

python main.py --input_n [number of historical sequence frames] --output_n [maximum number of predicted frames] --test_output_n [index of the test frame] --skip_rate [sampling rate] --n_pre [number of dct coefficients] --mode test --model_path ./checkpoints/CKPT_3D_H36M --global_translation --data_dir ./data --version [long / short]

Visualization

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

python main.py --input_n [number of historical sequence frames] --output_n [maximum number of predicted frames] --skip_rate [sampling rate] --n_pre [number of dct coefficients] --mode viz --model_path ./checkpoints/CKPT_3D_H36M --n_viz 5 --data_dir ./data --version [long / short]

With global translation, we have the following commands:

python main.py --input_n [number of historical sequence frames] --output_n [maximum number of predicted frames] --skip_rate [sampling rate] --n_pre [number of dct coefficients] --mode viz --model_path ./checkpoints/CKPT_3D_H36M --n_viz 5 --global_translation --data_dir ./data --version [long / short]

Acknowledgments

This code is based on the implementations of STSGCN.