Diverse Video Generation

This is code for the paper Diverse Video Generation using a Gaussian Process Trigger (ICLR 21) by Gaurav Shrivastava and Abhinav Shrivastava.

KTH action dataset

First download the KTH action recognition dataset by running:

sh data/download_kth.sh /my/kth/data/path/

where /my/kth/data/path/ is the directory the data will be downloaded into. Next, convert the downloaded .avi files into .png's for the data loader. To do this you'll want ffmpeg installed. The following script will do the conversion:

th data/convert_kth.lua --dataRoot /my/kth/data/path/ --imageSize 64

Experimental results in the paper used 64x64. To train the DVG model on 64x64 KTH videos run:

python train.py --data_root //path/to/the/data --image_width 64 --dataset kth --channels 1

To generate frames using a pretrained DVG model on 64x64 KTH videos run:

python generate_frames.py --data_root //path/to/the/data --image_width 64 --dataset kth --channels 1

To cite our work please use the following

@article{shrivastava2021diverse,
  title={Diverse Video Generation using a Gaussian Process Trigger},
  author={Shrivastava, Gaurav and Shrivastava, Abhinav},
  journal={arXiv preprint arXiv:2107.04619},
  year={2021}
}

@phdthesis{shrivastava2021diverse,
  title={Diverse Video Generation},
  author={Shrivastava, Gaurav},
  year={2021},
  school={University of Maryland, College Park}
}