Unsupervised light field video stabilization method based on sub-aperture pose optimization

Abstract

The application discloses an unsupervised light field video anti-shake method based on sub-aperture pose optimization. The application utilizes the feature representation of a sub-aperture rich light field image, and simultaneously takes the camera pose estimated by a central sub-aperture image as a benchmark to design an unsupervised loss function based on sub-aperture pose joint optimization. The sub-aperture image is used to learn the scene high-dimensional depth features and the offset matrix of the edge sub-aperture relative to the central sub-aperture to perform joint optimization to ensure the consistency of the stable motion of the sub-aperture, so that the sub-aperture picture motion consistent light field video anti-shake is realized. The unsupervised learning method solves the problem that the video anti-shake algorithm depends on a high-quality data set.

Images