Final Project for MIT 6.819 Advances in Computer Vision

pytorch
skimage
numpy
cv2
tqdm

TODO

This project is based on the following two papers.

Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network

https://arxiv.org/abs/1609.04802

Frame-Recurrent Video Super-Resolution

https://arxiv.org/abs/1801.04590

We are trying to combine these two models together in our project (this is the where the name come from "Frame Recurrent Super Resolution GAN").

Our final report is available here: Report

We used the dataset from toflow-dataset, which is about 15GB, containing 7.8k video clips with 7 frames per clip.

We trained our model on the dataset for 9 epochs. and compared with default SRGAN model in the repo. We did not retrain SRGAN in our dataset, so the result is for reference only.

The following results are produced by 7-epoch model.

We referred to SRGAN Implementation by LeftThomas in our project. We would like to thank the author for his fabulous work.