Cornell University Researchers in Artificial Intelligence propose a novel neural network framework to address the video mating problem
The most popular computer applications are video and image editing. Image and video editing has been studied in depth through neural network architectures since the advent of Machine Learning and Deep Learning. Until recently, the majority of Deep Learning (DL) models for video and image editing were supervised. More specifically, the training data had to include pairs of input data and output data that would be used to learn the details of desired transformation. Recently, end-toend learning frameworks that require only one image as input have been proposed.
Video matting is an important part of video editing. The term \”matting\” dates back to 19th century, when matte paint was painted on glass plates and placed in front of cameras during filming. This created the illusion that an environment wasn’t present at the location of the filming. The composition of digital images is similar today. The intensity of foregrounds and backgrounds of each image is shaded using a composite formula.
This process is very powerful but has limitations. This process requires a clear factorization of an image into layers foreground, background and then assumes that each layer can be treated independently. The layers decomposition can be a difficult task in some situations, such as video matting.