一种提升3维重建精度和效率的图像选择方法

Large scale 3D scene reconstructions are usually realized via merging local depth maps with heavy computational burden, so that the accuracy of local 3D data will affect the accuracy and the efficiency of scene reconstruction. To tackle this problem, in this paper, an image selection method to improve the computation efficiency and accuracy of 3D reconstruction was proposed. To improve the efficiency, a reference image selection method based on bit operation was proposed. Specially, a minimum calibrated images subset covering all three-dimensional points of the scene was selected as the reference image set, then an evaluation function of neighbor image selection was defined according to the scale evaluation factor, the angle between two camera main optical axes and the angle between 3D points and camera main optical axes. In implementation of the neighbor image selection algorithm, the method firstly removed most of the neighbor image candidates by the overlapping rate of reference image and candidates, and then sorted the remaining candidates by the evaluation indicator with descending order. Effective reference images selection can greatly reduce the amount of depth map calculation in multi-view 3D reconstruction of large scene, while effective adjacent images selection took into account the reconstruction accuracy and integrity. The proposed method was compared with another two methods and the COLMAP algorithm on two public data sets with ground truth and two data sets collected from real outdoor scenes. Experimental results demonstrated that the proposed method saved computing time by 53%～59%, improved the completeness by 4%～6% and the accuracy by 4%～7% in contrast to the other two image selection methods. Besides, the proposed method saved computing time more than 16%, improved the completeness by 3% and the accuracy about 2.7% in contrast to the state-of-the-art COLMAP method. Experimental results on multiple data sets demonstrated the effectiveness of the proposed method.