Enhanced Bi-directional Motion Estimation for Video Frame Interpolation

Introduction

This project is an implementation of our WACV 2023 paper, Enhanced Bi-directional Motion Estimation for Video Frame Interpolation.

We refer our frame interpolation method as EBME (Enhanded Bi-directional Motion Estimation for video frame interpolation). Our motion estimator is designed by carefully adapting the ingredients (e.g., warping, correlation) in optical flow research for simultaneous bi-directional motion estimation within a flexible pyramid recurrent framework. It is extremely lightweight (15x smaller than PWC-Net), yet enables reliable handling of large and complex motion cases. Based on estimated bi-directional motions, we employ a synthesis network to fuse forward-warped representations and predict the intermediate frame. Our method achieves excellent performance on a broad range of frame interpolation benchmarks.

Preparation

Python and Cuda environment

Our code has been tested with PyTorch 1.6 and Cuda 10.2. We believe that it is also compatible with higher version of PyTorch. CuPy package is required for running the forward warping operation (refer to softmax-splatting for details).

conda create --name ebme python=3.7
conda activate ebme
conda install pytorch==1.6.0 torchvision==0.7.0 cudatoolkit=10.2 -c pytorch
pip3 install cupy_cuda102==9.4.0
pip3 install -r requirements.txt

If your Cuda version is lower than 10.2 (not lower than 9.2), we suggest to replace cudatoolkit=10.2 in above command with cudatoolkit=9.2, and replace cupy_cuda102==9.4.0 with cupy_cuda92==9.6.0.

Trained model weights

We have placed our trained model weights in ./checkpoints. The weight of our base version of EBME model is placed in ./checkpoints/ebme, and the weight of our high-resolution version of model is placed in ./checkpionts/ebme-h.

Training and test datasets

We train our model on Vimeo90K dataset, and evaluate it on Vimeo90K, UCF101, SNU-FILM, and 4K1000FPS.

If you want to train and benchmark our model, please download Vimeo90K, UCF101, SNU-FILM, 4K1000FPS.

Play with demo

We provide a script to test our frame interpolation pipeline, given two consecutive input frames, and the desired time step. Run the following command, then you will obtain interpolated frame and bi-directional flows in the ./demo/output directory.

python3 -m demo.interp_imgs \
--frame0 demo/images/caption0.png \
--frame1 demo/images/caption1.png \
--time_period 0.5

Here the time_period (float number in 0~1) indicates the time step of the intermediate frame you want to interpolate.

Training on Vimeo90K

By default, our model is trained on Vimeo90K. We have set default hyper-parameters for training, except for the path of training dataset. You can run the following command to train our base version of EBME:

CUDA_VISIBLE_DEVICES=0,1 python3 -m torch.distributed.launch \
    --nproc_per_node=2 --master_port=10000 -m tools.train \
        --world_size=2 \
        --data_root /path/to/vimeo_triplet \
        --train_log_root /path/to/train_log \
        --exp_name ebme-baseline \
        --batch_size 16 \
        --nr_data_worker 2

Please assign the root dir to save logs. Otherwise, all training logs will be saved in ../ebme-train-log by default.

If you want to train our high-resolution version of EBME, we suggest to use 4 GPUs for training, and slightly change some hyper-parameters.

CUDA_VISIBLE_DEVICES=0,1,2,3 python3 -m torch.distributed.launch \
    --nproc_per_node=4 --master_port=10000 -m tools.train \
        --world_size=4 \
        --data_root /path/to/vimeo_triplet \
        --train_log_root /path/to/train_log \
        --exp_name ebme-high-resolution \
        --high_synthesis True \
        --batch_size 8 \
        --nr_data_worker 4

All data generated during the training process (including the trained models) will be saved in ./train-log directory. We intentionally soft link ../embe-train-log to ./train-log, in order to avoid saving files under dir of this codebase.

Benchmarking

Benchmarking scripts

We provide scripts to test frame interpolation accuracy on Vimeo90K, UCF101, SNU-FILM, and 4K1000FPS. You should configure the path to benchmark datasets when running these scripts.

python3 -m tools.benchmark_vimeo90k --data_root /path/to/vimeo_triplet/
python3 -m tools.benchmark_ucf101 --data_root /path/to/ucf101/
python3 -m tools.benchmark_snufilm --data_root /path/to/SNU-FILM/
python3 -m tools.benchmark_8x_4k1000fps --test_data_path /path/to/4k1000fps/test

In order to benchmark our high-resolution version of EBME, you need to change the path of model file, and set high_synthesis as True.

Additionally, run the following command can test our runtime.

python -m tools.runtime

Our benchmarking results

Our benchmarking results on UCF101, Vimeo90K, SNU-FILM are shown in below table. You can verify our results by running our benchmarking scripts. Runtime is measured with a single 2080TI GPU for interpolating two 640x480 frames.

Our benchmarking results on 4K1000FPS are shown in below table.

Acknowledgement

We borrow some codes from RIFE and softmax-splatting. We thank the authors for their excellent work. When using our code, please also pay attention to the license of RIFE and softmax-splatting.

Citation

@inproceedings{jin2023enhanced,
  title={Enhanced Bi-directional Motion Estimation for Video Frame Interpolation},
  author={Jin, Xin and Wu, Longhai and Shen, Guotao and Chen, Youxin and Chen, Jie and Koo, Jayoon and Hahm, Cheul-hee},
  booktitle={Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
  year={2023}
}