Abstract
Camera calibration is a fundamental problem of computer vision. It is vital both for geometric measurements and image processing techniques like image stitching. Traditional camera calibration procedures do not take into account local irregularities in the mapping. They rather describe the distortions using global models for the whole camera target. Approaches to measure the camera distortions independently using dynamic calibration patterns displayed on a monitor screen can overcome this restriction. However, their accuracy is limited by using bi-level patterns. In this article, the utilization of continuous-tone calibration patterns is enabled by applying a per-pixel photometric calibration of the monitor-camera response function. Three different calibration methods are described and compared. The best performer, the Empirical Model of Response, achieves an overall calibration RMSE of 0.893 brightness levels. This photometric calibration is applied to calibration patterns having a Gaussian modulated brightness, which are used to perform position estimation in order to measure geometric distortions. These patterns clearly outperform bi-level patterns as their estimation RMSE is 0.0092 pixels, which is 5.13 times more accurate than an estimation based on bi-level patterns. The performance of camera calibration procedures based on position estimation will be enhanced by continuous-tone patterns.