Multiview 3D reconstruction based on vanishing points and homography

Abstract

3D reconstruction is one of the long lasting research topics in computer vision. It is the task of regenerating a model in three-dimensional (3D) space by images taken from the scene. An image-based method for 3D reconstruction based on intersection of planar visual hulls from projective images taken by multiview videos is proposed. Reconstruction can be performed without any further information related to camera calibration, feature extraction, and dense matching. This process computes the homography of reference plane along with one corresponding point off the reference plane, and vanishing points in all views. The 2D visual hulls along the object structure are generated. By stacking 2D visual hull slices, the 3D model of the object is obtained efficiently with least imposed constraints.

Full Explicit Consistency Constraints in Uncalibrated Multiple Homography Estimation

“…with J ML given in (22) and Π i (η) given in (4), was optimised using the Levenberg-Marquardt algorithm. To optimise J ML subject to the explicit constraint (21), we used MATLAB's fmincon function set to the interior point algorithm.…”

“…A task closely related to estimating a single homography is that of estimating multiple homographies. Multiple planar surfaces are ubiquitous in urban environments, and, as a result, estimating multiple homographies 1 arXiv:1805.02352v6 [cs.CV] 19 Feb 2019 between two views from image measurements is an important step in many applications such as non-rigid motion detection [21,40], enhanced image warping [16], multiview 3D reconstruction [22], augmented reality [30], indoor navigation [32], multi-camera calibration [37], camera-projector calibration [28], or ground-plane recognition for object detection and tracking [1]. Surprising as it may seem, a vast array of techniques for estimating multiple homographies, including many robust multi-structure estimation methods [5,6,12,15,20,25,29,38,39,42] applicable to the task of estimating multiple homographies, are deficient in a fundamental waythey fail to recognise that a set of homography matrices does not represent a set of genuine homographies between two views of the same scene unless appropriate consistency constraints are satisfied.…”