shuyueL / FFNet

Implementation of our CVPR 2018 Paper (FFNet: Video Fast-Forwarding via Reinforcement Learning)

FFNet

This repository provides the implementation for video fast-forward with reinforcement learning, i.e. FFNet in our paper:

FFNet: Video Fast-Forwarding via Reinforcement Learning

If you find the codes or other related resources from this repository useful, please cite the following paper:

@inproceedings{lan2018ffnet,
  title={FFNet: Video Fast-Forwarding via Reinforcement Learning},
  author={Lan, Shuyue and Panda, Rameswar and Zhu, Qi and Roy-Chowdhury, Amit K},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
  pages={6771--6780},
  year={2018}
}

Environment

• Windows or Linux
• NVIDIA GPU with compute capability 3.5+
• Python 3.5
• Tensorflow

Data

The original data we used in paper are available from the following websites

• Tour20 dataset: https://vcg.ece.ucr.edu/
• TVSum dataset: https://github.com/yalesong/tvsum

Codes

Testing

We offer a testing example with a pre-trained model in the ./model directory. Download this repository and run the following command:

python nn_test.py

The fast-forward result will be in the ./output directory.

Training

If you want to train the model on your own data, you can find the script for training in nn_train.py. For more details, please refer to our paper.