EXTRACTING CAMERA POSE USING SINGLE IMAGE SUPER RESOLUTION NETWORKS

This work proposes a mechanism which can be used as a basis for allowing camera POSE information to be maintained reliably during loss or interference with inertial motion unit or positioning system integration. This basis is formed by employing image synthesis networks with atypical data for the network type: inputs are normal down scaled source imagery while outputs are native resolution images composed of the contents of the same scene viewed from a fixed offset position. The goal of this application is to simulate the presence of a binary camera from monocular hardware, which makes feasible certain POSE estimation workflows which would normally require binary cameras on monocular platforms. Being able to rapidly synthesize images of additional camera positions without having to physically navigate to those positions allows for two methods to build off each other. First, knowing that the model should consistently maintain a specific POSE from the source camera allows synthetic images to be used to artificially inflate available data during structure from motion processing with confidence in the accuracy of synthetic points. It also enables the comparison of an image at an actual physical location with the synthetic one later as a measure of POSE accuracy which can be incorporated into a solution for computing POSE of the image source.