The code in this repo is used to take part in the video repair contest, whose task is to remove artifacts in video caused by JPEG compression. You can click this link to see more details.
The code is tested under Python 3.5.2, Tensorflow 1.8.0, CUDA 10 and cuDNN 7.1.4.
Run make
to compile deform_conv ops.
The majority of code is modified from EDVR. What I do is to find a more appropriate reconstruction model for video repair. The model is shown as follows.
Firstly, you can have a look at scripts under tools folder to prepare data used to train.
Secondly, configure params.py
to make sure the code will work as you desire.
Finally, run python train.py
to train your own model.
deform_conv ops is copied from TF-deformable-conv.