Implementing Adaptive Separable Convolution for Video Frame Interpolation

September 21, 2018 · View on GitHub

This is a fully functional implementation of the work of Niklaus et al. [1] on Adaptive Separable Convolution, which claims high quality results on the video frame interpolation task. We apply the same network structure trained on a smaller dataset and experiment with various different loss functions, in order to determine the optimal approach in data-scarce scenarios.

For detailed information, please see our report on arXiv:1809.07759.

The video below is an example of the capabilities of this implementation. Our pretrained model (used in this instance) can be downloaded from here.

Installation

Note that the following instructions apply in the case of a fresh Ubuntu 17 machine with a CUDA-enabled GPU. In other scenarios (ex. if you want to work in a virtual environment or prefer to use the CPU), you may need to skip or change some of the commands.

Install pip3:

sudo apt-get install python3-setuptools
sudo easy_install3 pip

Install GCC & friends:

sudo apt-get install gcc libdpkg-perl python3-dev

Install FFmpeg:

sudo apt-get install ffmpeg

Install CUDA:

curl -O https://developer.download.nvidia.com/compute/cuda/repos/ubuntu1604/x86_64/cuda-repo-ubuntu1604_8.0.61-1_amd64.deb
sudo dpkg -i ./cuda-repo-ubuntu1604_8.0.61-1_amd64.deb
sudo apt-get update
sudo apt-get install cuda-8-0 -y

Install NVCC:

sudo apt-get install nvidia-cuda-toolkit

Install dependencies:

sudo pip3 install -r ./path/to/requirements.txt

How to setup the project and train the network:

Move to the project directory:

cd ./sepconv

Create a new configuration file:

echo -e "from src.default_config import *\r\n\r\n# ...custom constants here" > ./src/config.py

To train the network, run `main.py` as a module:

python3 -m src.main