This project aims to build a model that takes a blurry image as input (e.g. due to an object while being photographed) and infers what that image would look like if it weren't blurred.
The data used for training consists of unlabeled videos. I create the blurred images by overlaying subsequent frames from the video, and I take a single frame from that sequence as the unblurred image.
The project is in a sufficiently early stage that this repository should be viewed as nothing but a space for scratch work.