Refining single view calibration with the aid of metric scene

Abstract

Intrinsic camera calibration using a single image is possible provided that certain geometric objects such as orthogonal vanishing points and metric homographies can be estimated from the image and give rise to adequate constraints on the sought calibration parameters. In doing so, however, any additional metric information that might be available for the imaged scene is not alwaysl traightforward to accommodate. This paper puts forward a method for incorporating into the calibration procedure metric scene information expressed in the form of known segment 3D angles, equal but unknown 3D angles and known 3D length ratios.l Assuming the availability of an initial calibration estimate, the proposed method refines the former by numerically minimizing 4n error term corresponding to the discrepancy between the scene's known metric properties and the values measured with the laid of the calibration estimate. Sample experimental results demonstrate the improvements in the intrinsic calibration estimates that are achieved by the proposed method.