Line-Based Geometric Consensus Rectification and Calibration From Single Distorted Manhattan Image

Recent advances in single image rectification and intrinsic calibration has been addressed by employing line information on the distorted image. The core issues of this technique are the separation of rectification and calibration procedures, and the suffering of geometric nonconformity. In this work, we propose a novel Geometric Consensus Rectification and Calibration algorithm, which we refer to as GCRC framework. We show how the geometric consensus rectification and calibration can be performed in a unified framework and solve the above issues. The proposed GCRC not only guarantees the geometrical consensus on the rectified images, but allows us to perform the robust intrinsic parameters estimation with the grouped circular arcs. Through “grouping by voting” in a unified framework, the geometric consensus rectification and calibration are robustly conducted on single distorted Manhattan images. Experiments on a number of distorted images, including the simulated YorkUrbanDB dataset, Panoramic Fisheye dataset, checkerboard image, and Internet images, demonstrate that the GCRC significantly improve the performance of geometrically consensus rectification and intrinsic parameters estimation. In particular, the GCRC shows relatively small variations with a different number of lines, which outperforms various previous approaches.