Multi-camera Array Calibration for Light Field Depth Estimation

At the core of stereo methods for depth estimation and 3D reconstruction lies geometric calibration, i.e. the determination of intrinsic and extrinsic camera parameters and consecutive image rectification, such that the epipolar constraints are met in all views. In this spotlight paper, we present a multi-camera array calibration that fulfills the requirements for 3D reconstruction. The method is based on an optimization procedure that minimizes the reprojection error. We used it to calibrate the Xapt Eye-sect XA camera array with 4x4 camera modules equipped with identical wide-angle lenses. For this particular setup, we analyzed the algorithm’s precision step by step, from initial pairwise multi-view stereo calibration to final bundle adjustment, to assess influence of each individual step. The conducted quantitative analysis based on the reprojection error revealed superiority of the bundle adjustment over all other considered intermediate steps yielding accuracy as much as 33x higher than the initial pairwise method. In order to demonstrate real-world performance of the calibrated camera array, we present a number of acquisitions of different physical objects along with estimated disparity maps and corresponding texture images generated by a light field multi-view stereo algorithm.