FILM: Frame Interpolation for Large Scene Motion

Project | Paper | YouTube | Benchmark Scores

Tensorflow 2 implementation of our high quality frame interpolation neural network. We present a unified single-network approach that doesn't use additional pre-trained networks, like optical flow or depth, and yet achieve state-of-the-art results. We use a multi-scale feature extractor that shares the same convolution weights across the scales. Our model is trainable from frame triplets alone.

FILM: Frame Interpolation for Large Motion
Fitsum Reda, Janne Kontkanen, Eric Tabellion, Deqing Sun, Caroline Pantofaru, Brian Curless
Google Research
Technical Report 2022.

FILM transforms near-duplicate photos into a slow motion footage that look like it is shot with a video camera.

Installation

• Get Frame Interpolation source codes

> git clone https://github.com/google-research/frame-interpolation frame_interpolation

• Optionally, pull the recommended Docker base image

> docker pull gcr.io/deeplearning-platform-release/tf2-gpu.2-6:latest

• Install dependencies

> pip install -r frame_interpolation/requirements.txt
> apt-get install ffmpeg

Pre-trained Models

• Create a directory where you can keep large files. Ideally, not in this directory.

> mkdir 
   

   

• Download pre-trained TF2 Saved Models from google drive and put into .

The downloaded folder should have the following structure:

pretrained_models/
├── film_net/
│   ├── L1/
│   ├── VGG/
│   ├── Style/
├── vgg/
│   ├── imagenet-vgg-verydeep-19.mat

Running the Codes

The following instructions run the interpolator on the photos provided in frame_interpolation/photos.

One mid-frame interpolation

To generate an intermediate photo from the input near-duplicate photos, simply run:

> python3 -m frame_interpolation.eval.interpolator_test \
     --frame1 frame_interpolation/photos/one.png \
     --frame2 frame_interpolation/photos/two.png \
     --model_path 
   
    /film_net/Style/saved_model \
     --output_frame frame_interpolation/photos/middle.png \

   

This will produce the sub-frame at t=0.5 and save as 'frame_interpolation/photos/middle.png'.

Many in-between frames interpolation

Takes in a set of directories identified by a glob (--pattern). Each directory is expected to contain at least two input frames, with each contiguous frame pair treated as an input to generate in-between frames.

> python3 -m frame_interpolation.eval.interpolator_cli \
     --pattern "frame_interpolation/photos" \
     --model_path 
   
    /film_net/Style/saved_model \
     --times_to_interpolate 6 \
     --output_video

   

You will find the interpolated frames (including the input frames) in 'frame_interpolation/photos/interpolated_frames/', and the interpolated video at 'frame_interpolation/photos/interpolated.mp4'.

The number of frames is determined by --times_to_interpolate, which controls the number of times the frame interpolator is invoked. When the number of frames in a directory is 2, the number of output frames will be 2^times_to_interpolate+1.

Datasets

We use Vimeo-90K as our main training dataset. For quantitative evaluations, we rely on commonly used benchmark datasets, specifically:

• Vimeo-90K
• Middlebury-Other
• UCF101
• Xiph

Creating a TFRecord

The training and benchmark evaluation scripts expect the frame triplets in the TFRecord storage format.

We have included scripts that encode the relevant frame triplets into a tf.train.Example data format, and export to a TFRecord file.

You can use the commands python3 -m frame_interpolation.datasets.create_ _tfrecord --help for more information.

For example, run the command below to create a TFRecord for the Middlebury-other dataset. Download the images and point --input_dir to the unzipped folder path.

> python3 -m frame_interpolation.datasets.create_middlebury_tfrecord \
    --input_dir=
   
     \
    --output_tfrecord_filepath=
    


   

Training

Below are our training gin configuration files for the different loss function:

frame_interpolation/training/
├── config/
│   ├── film_net-L1.gin
│   ├── film_net-VGG.gin
│   ├── film_net-Style.gin

To launch a training, simply pass the configuration filepath to the desired experiment.
By default, it uses all visible GPUs for training. To debug or train on a CPU, append --mode cpu.

> python3 -m frame_interpolation.training.train \
     --gin_config frame_interpolation/training/config/
   
    .gin \
     --base_folder 
     \
     --label 
    

    
   

• When training finishes, the folder structure will look like this:

   
    /
├── 
    /
│   ├── config.gin
│   ├── eval/
│   ├── train/
│   ├── saved_model/

   

Build a SavedModel

Optionally, to build a SavedModel format from a trained checkpoints folder, you can use this command:

> python3 -m frame_interpolation.training.build_saved_model_cli \
     --base_folder  \
     --label 
   

   

• By default, a SavedModel is created when the training loop ends, and it will be saved at / / .

Evaluation on Benchmarks

Below, we provided the evaluation gin configuration files for the benchmarks we have considered:

frame_interpolation/eval/
├── config/
│   ├── middlebury.gin
│   ├── ucf101.gin
│   ├── vimeo_90K.gin
│   ├── xiph_2K.gin
│   ├── xiph_4K.gin

To run an evaluation, simply pass the configuration file of the desired evaluation dataset.
If a GPU is visible, it runs on it.

> python3 -m frame_interpolation.eval.eval_cli -- \
     --gin_config frame_interpolation/eval/config/
   
    .gin \
     --model_path 
    
     /film_net/L1/saved_model

    
   

The above command will produce the PSNR and SSIM scores presented in the paper.

Citation

If you find this implementation useful in your works, please acknowledge it appropriately by citing:

@inproceedings{reda2022film,
 title = {Frame Interpolation for Large Motion},
 author = {Fitsum Reda and Janne Kontkanen and Eric Tabellion and Deqing Sun and Caroline Pantofaru and Brian Curless},
 booktitle = {arXiv},
 year = {2022}
}

@misc{film-tf,
  title = {Tensorflow 2 Implementation of "FILM: Frame Interpolation for Large Scene Motion"},
  author = {Fitsum Reda and Janne Kontkanen and Eric Tabellion and Deqing Sun and Caroline Pantofaru and Brian Curless},
  year = {2022},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/google-research/frame-interpolation}}
}

Contact: Fitsum Reda ([email protected])

Acknowledgments

We would like to thank Richard Tucker, Jason Lai and David Minnen. We would also like to thank Jamie Aspinall for the imagery included in this repository.

Coding style

• 2 spaces for indentation
• 80 character line length
• PEP8 formatting

Disclaimer

This is not an officially supported Google product.