Deep Single Fisheye Image Camera Calibration for Over 180-degree Projection of Field of View

We propose a learning-based calibration method for trigonometric function models that represent distortion with over 180° projection of field of views. Unlike perspective projection for less than 180° projection of field of views, fisheye projection such as equisolid angle projection is valid for whole world coordinates. To calibrate fisheye camera models, we define a new loss function based on camera projection effectively to optimize fisheye camera extrinsic (tilt and roll angles) and intrinsic (focal length) parameters. Our loss achieves small prediction errors throughout the ranges of parameters. Our results show that our method predicts precise fisheye camera parameters compared with conventional polynomial function models for radial distortion. This work is the first to calibrate a fisheye camera model including extrinsic and intrinsic parameters for over 180° projection of field of views from a single image to our knowledge.