A Technique for Video Inpainting using Deep Learning

With the advent of internet and social media, the number of videos being shot on a daily basis has increased manifold. However, some of these videos don't turn out the way you want them to. Video Inpainting is a process which can be used to eliminate obvious flaws in such videos and make them better. Further, we see a ton of highly enthralling videos in action movies, especially superhero movies. Most of the action sequences are not real, none of the cameramen will be able to shoot it to perfection in such unreal scenarios and surroundings. All these videos are doctored using a variety of video inpainting methods. In our paper, we present one such Video Inpainting method built using several deep learning techniques. Our Video Inpainting network uses a set of reference frames to regenerate the holes in a given target frame. It establishes suitable relationships between the reference frames and the target frame, stores them in an asymmetric key value store and goes about the process of inpainting. We validate that our model produces visually pleasing videos in all realistic scenarios.