一种提升3维重建精度和效率的图像选择方法

An Image Selection Method for Efficient and Accurate 3D Scene Reconstruction

大型场景3维重建常需融合各局部深度图,故局部3维数据精确度直接关系整个场景重建精确度,且计算量巨大。针对此问题,提出一种旨在提高3维计算效率和精度的图像选择方法。为提高计算效率,提出一种基于位运算的参考图像选择方法,选取可覆盖场景全部3维点的最小标定图像集子集作为参考图像集;根据影响重建精度的尺度评价因子、参考与邻近图像对应相机主光轴夹角及3维点与相机主光轴夹角建立邻近图像选择的评价函数;在邻近图像的具体选择上,先根据参考图像与邻近图像的覆盖率缩小搜索范围,再根据评价函数对剩余图像进行降序排序,并选择前3张图像作为该参考图像的邻近图像。有效的参考图像选择能够极大地减少大场景多视立体3维重建过程中深度图计算量,有效的邻近图像选择兼顾了重建结果的精度和完整度。将本文方法在两组具有真实3维点云数据的公开数据集和两组室外真实场景拍摄的数据集上与另外两种选图方法以及COLMAP方法进行了对比。实验结果表明:本文方法相比于另外两种选图算法时间节省53%~59%,完整性提高4%~6%,精度提高4%~7%;相比于先进的COLMAP算法时间节省16%以上,完整性提高3%,精度提高2.7%左右。实验结果证明了本文方法的有效性。

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.