Self-Teaching: Unsupervised Learning of Optical Flow with Non-occlusion and Full-Image Wrapping from Geometry

Abstract

This paper presents a self-teaching algorithm for learning optical flow with non-occlusion and full-image wrapping from geometry. The proposed algorithm uses a monocular camera in order to learn the optimal state-of-the-art optical flow. It is based on a novel approach which uses the camera calibration parameters to directly optimize the flow. In this way, the algorithm is able to learn the full image-space motion from a single image pair. The proposed approach is evaluated on several datasets and is shown to be competitive with state-of-the-art supervised learning methods. The main benefits of the proposed approach is that it does not require large datasets for training and is able to leverage the natural geometric constraints of the scene to obtain the best possible results. The proposed algorithm is also able to produce more spatially consistent optical flow than existing deep learning approaches.