Efficient large-scale geometric verification for structure from motion
Efficient large-scale geometric verification for structure from motion
复制标题
DOI：
10.1016/j.patrec.2018.09.028
复制
发表时间：
2019-07
期刊：
影响因子：
--
通讯作者：
Qingshan Xu;Jie-gu Li;Wenbing Tao;Delie Ming
中科院分区：
文献类型：
--
作者：
Qingshan Xu;Jie-gu Li;Wenbing Tao;Delie Ming
几何验证是极线几何中的一个基本问题，它通过估计基本矩阵和单应矩阵来确定图像对是否具有共同的三维结构。然而，它仍然遭受低效率时，遇到大规模的结构从运动（SfM）。本文采用线性同余算法对点对进行并行采样。然后，我们建议在GPU中同时估计一定数量的候选基本/单应性矩阵，以避免迭代随机点集采样，并基于这些点集执行矩阵估计，然后进行最佳矩阵选择和进一步并行细化。在大量数据集上的实验表明，我们基于GPU的几何验证比原始迭代方法快70倍，同时保持相当满意的3D重建结果。
Geometric verification is a fundamental problem in epipolar geometry, which estimates fundamental matrix and homography matrix to confirm that image pairs share a common 3D structure. However, it still suffers from low efficiency when it encounters large-scale Structure from Motion (SfM). In this paper, we adopt the linear congruence algorithm to sample point-pairs in parallel. Then, we propose to simultaneously estimate a certain number of candidate fundamental/homography matrices in GPU to avoid the iterative random point sets sampling and perform matrix estimation based on these point sets, followed by best matrix selection and further parallel refinement. Experiments on extensive datasets show that our GPU-based geometric verification is up to seventy times faster than original iteration method while maintaining comparable satisfactory 3D reconstruction results.