Code repository for BMVC paper LARNet: Latent Action Representation for Human Action Synthesis Paper available at [arxiv] (https://arxiv.org/pdf/2110.10899.pdf)
Project page with more information available in [CRCV webpage] (https://www.crcv.ucf.edu/research/projects/larnet-latent-action-representation-for-human-action-synthesis)
Demo available HERE
This is an implementation of LARNet on NTU-RGB+D 60 dataset. It is built using the PyTorch library. The datasets have to be downloaded separately.
Please install following prerequisites: Python>=3.6, Pytorch>=1.6
Install remainig libraries from requirements.txt using
pip install -r requirements.txt
Download datasets and set the data_root_dir variable in train.py for each dataset.
Download I3D pretrained weights for Charades / Kinetics and setup the path in i3d_weights_path variable in train.py.
If you are resuming training, set the weights for generator and discriminator and uncomment weight load command for each in train.py.
After doing setup and putting data/model path correctly in train.py, run using:
python train.py
If you enable tensorboard, it will save the generated video for each snapshot interval in the tensorboard log file.