A Deep Ordinal Distortion Estimation Approach for Distortion Rectification

Radial distortion has widely existed in the images captured by popular wide-angle cameras and fisheye cameras. Despite long history of rectification, accurately estimating parameters from a single distorted image is still challenging. The main reason that these are implicit to features, influencing networks learn information fully. In this work, we propose novel rectification approach can obtain more accurate with higher efficiency. Our key insight be cast as problem learning an ordinal image. To solve problem, design local-global associated estimation network learns approximate realistic distribution. contrast parameters, proposed explicit relationship significantly boosts perception neural networks. Considering redundancy information, our only uses patch for estimation, showing promising applications efficient rectification. field, first unify heterogeneous into learning-friendly intermediate representation through distortion, bridging gap between feature experimental results demonstrate outperforms state-of-the-art methods significant margin, approximately 23% improvement on quantitative evaluation while displaying best performance visual appearance.