Convolutional Neural Networks for Camera Self-Calibration With Varying Intrinsic Parameters

Abstract

Stereo camera self-calibration is a complex problem that has been tackled by computer vision and pattern recognition communities, and significant progress has been made. In the same context, this work proposes a new and robust fully automated approach of camera self-calibration with varying intrinsic parameters from two images of an unknown 3D object. This method is based on deep learning, such as convolutional neural networks (CNN), to reduce mathematical complexities and estimation time to a minimum. These networks use a special architecture that is particularly well adapted to estimate the intrinsic parameters of the camera through back propagation and gradient descent. First, our learningbased method introduces the self-calibration equations formulated into a non-linear cost function. The new optimization of this cost function makes it possible to estimate the intrinsic parameters of the cameras. The strong point of the proposed architecture is that it merges the high effectiveness of the CNN algorithm with the four pillars of our non-linear cost function formulation (two images, four matches, any camera and unknown 3D scene) to form the fast and robust proposed network. The experimental results on synthetic and real data prove the performance of the proposed technique in terms of simplicity, precision, and convergence.