FFCVSR (AAAI 2019)

AAAI 2019 paper "Frame and Feature-Context Video Super-Resolution" [1]
Paper

FFCVSR-motion is a improved version for FFCVSR, which adds motion prediction, feature alignment and gate selection. The new version paper is submited to TPAMI 2020 and under review.

Code

We release FFCVSR and FFCVSR-motion inference model and FFCVSR-motion training code in REDS dataset.

Our testing environment is:

• TensorFlow == 1.9
• Python 3.6
• NVIDIA GTX 1080Ti

Inference

1. Download the pretrained checkpoints from WeiYun: https://share.weiyun.com/sEHySs5d

2. Testing model in VID4 dataset

# testing FFCVSR model
python test_VID4_FFCVSR.py

# testing FFCVSR-motion model
python test_VID4_FFCVSR_motion.py

# compile inverse_warp cuda verison to speed up the FFCVSR-motion model if the OS is linux
cd custom_op
make

Training

1. Download the REDS dataset (sharp type): https://seungjunnah.github.io/Datasets/reds

2. Put the REDS dataset in datasets/REDS

3. Generate tfrecords for REDS:

python tfrecords/gen_REDS_tfrecords.py

4. Train the FFCVSR-motion

python train_REDS_FFCVSR_motion.py

VID4 Dataset Performance

Methods	Training Dataset	PSNR	SSIM	Inference Time
FFCVSR	Internet Videos	26.97	0.815	28.4 ms
FFCVSR-motion	REDS sharp	27.15	0.821	38.6 ms

Citation

[1]  @inproceedings{ffcvsr,
         author = {Bo Yan, Chuming Lin, and Weimin Tan},
         title = {Frame and Feature-Context Video Super-Resolution},
         booktitle = {AAAI},
         year = {2019}
     }