A precise model of lens distortion is required in both Computer Vision for 3D scene analysis and in Computer Graphics, e.g. for placing rendered objects into camera-recorded scenes, where the rendering must match the lens characteristics of the camera.
The standard model for camera calibration assumes uniform lens distortion effects. This is problematic for low-cost lenses, e.g. for mobile applications, that may exhibit uneven distortions. Such lenses require a more fine-grained calibration that allows variations in distortion effects across the lens.
In this project, participating students will develop a lens calibration scheme that does not rely on a uniform distortion model and allows for lens irregularities.
For this purpose, a set of recorded images of known calibration targets (e.g. a checkerboard pattern) will be compared with images simulated using a rendering process that is differentiable with regard to the calibration parameters. This allows to apply optimization techniques for iterative refinement of these parameters.