Improving Camera Calibration for 3D Geometry Reconstruction using 3D Calibration Targets and Generalized Camera Models

In the field of 3D object reconstruction, camera calibration is the most essential step as it determines the accuracy of the final reconstruction. While this problem has seen many solutions of different kinds, we will develop a framework most suitable specifically for object reconstruction by using a calibrated camera to predict scene rays. In this work we will develop a new feature detection algorithm based on ring features and build up calibration targets with it. We create a new camera model that can handle arbitrary distortion. Additionally, we evaluate all methods using new metrics based on predicted scene rays. We will see that a generalized, more flexible distortion model adapts much better to the physical camera distortion and yields better calibration results. Moreover, we will see how the reprojection error can be misleading when evaluating a model for accuracy in 3D. In these cases, the newly devised metrics give a more holistic perspective. And finally we show how to effectively calibrate with non-planar calibration targets as effortlessly as with regular planar targets.