Awesome

EDSC-pytorch

Code for Multiple Video Frame Interpolation via Enhanced Deformable Separable Convolution [arXiv] .

Pre-trained models

Google Drive

Baidu Cloud : bdfu

Environment

We are good in the environment:

python 3.7

CUDA 10.1

Pytorch 1.0.0

opencv-python 4.2.0

numpy 1.18.1

cupy 6.0.0

Usage

We provide two versions of our model. The EDSC_s model was trained to generate the midpoint (in time) of the two input frames. And you can either choose the l1 or the lf model for distortion and perceptual quality, respectively.

We are good to run

python run.py --model EDSC_s --model_state EDSC_s_l1.ckpt --out out.png

The EDSC_m model is able to generate a frame at an arbirary time position. For instance, to generate an intermediate frame at t=0.1, we are good to run

python run.py --model EDSC_m --model_state EDSC_m.ckpt --time 0.1 --out out.png

Please see the paper for more details.

Citation

@article{EDSC,
    author={Cheng, Xianhang and Chen, Zhenzhong},
    journal={IEEE Transactions on Pattern Analysis and Machine Intelligence}, 
    title={Multiple Video Frame Interpolation via Enhanced Deformable Separable Convolution}, 
    year={2021},
    doi={10.1109/TPAMI.2021.3100714}
}

Acknowledgement

Part of the code was adapted from sepconv-slomo. A huge thanks to the authors!