Awesome

Channel Attention Is All You Need for Video Frame Interpolation

Myungsub Choi, Heewon Kim, Bohyung Han, Ning Xu, Kyoung Mu Lee

2nd place in [AIM 2019 ICCV Workshop] - Video Temporal Super-Resolution Challenge

Project | Paper-AAAI (Download the paper [here] in case the AAAI link is broken) | Poster

<a href="https://aaai.org/Papers/AAAI/2020GB/AAAI-ChoiM.4773.pdf" rel="Video"><img src="./figures/CAIN_paper_thumb.jpg" alt="Paper" width="75%"></a> <a href="https://www.dropbox.com/s/7lxwka16qkuacvh/AAAI-ChoiM.4773.pdf" rel="Poster"><embed src="./figures/CAIN_poster.pdf" alt="Poster" width="20%"></a>

Directory Structure

project
│   README.md
|   run.sh - main script to train CAIN model
|   run_noca.sh - script to train CAIN_NoCA model
|   test_custom.sh - script to run interpolation on custom dataset
|   eval.sh - script to evaluate on SNU-FILM benchmark
|   main.py - main file to run train/val
|   config.py - check & change training/testing configurations here
|   loss.py - defines different loss functions
|   utils.py - misc.
└───model
│   │   common.py
│   │   cain.py - main model
|   |   cain_noca.py - model without channel attention
|   |   cain_encdec.py - model with additional encoder-decoder
└───data - implements dataloaders for each dataset
│   |   vimeo90k.py - main training / testing dataset
|   |   video.py - custom data for testing
│   └───symbolic links to each dataset
|       | ...

Dependencies

Current version is tested on:

• Ubuntu 18.04
• Python==3.7.5
• numpy==1.17
• PyTorch==1.3.1, torchvision==0.4.2, cudatoolkit==10.1
• tensorboard==2.0.0 (If you want training logs)
• opencv==3.4.2
• tqdm==4.39.0

# Easy installation (using Anaconda environment)
conda create -n cain
conda activate cain
conda install python=3.7
conda install pip numpy
conda install pytorch torchvision cudatoolkit=10.1 -c pytorch
conda install tqdm opencv tensorboard

Model

Dataset Preparation

• We use Vimeo90K Triplet dataset for training + testing.
  • After downloading the full dataset, make symbolic links in data/ folder :
    • ln -s /path/to/vimeo_triplet_data/ ./data/vimeo_triplet
  • Then you're done!
• For more thorough evaluation, we built SNU-FILM (SNU Frame Interpolation with Large Motion) benchmark.
  • Download links can be found in the project page.
  • Also make symbolic links after download :
    • ln -s /path/to/SNU-FILM_data/ ./data/SNU-FILM
  • Done!

Usage

Training / Testing with Vimeo90K dataset

• First make symbolic links in data/ folder : ln -s /path/to/vimeo_triplet_data/ ./data/vimeo_triplet
  • Vimeo90K dataset
• For training: CUDA_VISIBLE_DEVICES=0 python main.py --exp_name EXPNAME --batch_size 16 --test_batch_size 16 --dataset vimeo90k --model cain --loss 1*L1 --max_epoch 200 --lr 0.0002
• Or, just run ./run.sh
• For testing performance on Vimeo90K dataset, just add --mode test option
• For testing on SNU-FILM dataset, run ./eval.sh
  • Testing mode (choose from ['easy', 'medium', 'hard', 'extreme']) can be modified by changing --test_mode option in eval.sh.

Interpolating with custom video

• Download pretrained models from [Here]
• Prepare frame sequences in data/frame_seq
• run test_custom.sh

Results

Video

<a href="https://www.dropbox.com/s/es2geeg26wwgezk/demo_combined_hires.mp4" rel="Video"><img src="./figures/CAIN_spotlight_thumb.jpg" alt="Video" width="100%"></a>

Citation

If you find this code useful for your research, please consider citing the following paper:

@inproceedings{choi2020cain,
    author = {Choi, Myungsub and Kim, Heewon and Han, Bohyung and Xu, Ning and Lee, Kyoung Mu},
    title = {Channel Attention Is All You Need for Video Frame Interpolation},
    booktitle = {AAAI},
    year = {2020}
}

Acknowledgement

Many parts of this code is adapted from:

• EDSR-Pytorch
• RCAN

We thank the authors for sharing codes for their great works.