分区优化混合SfM方法

针对大规模无序影像稀疏三维重建问题，本文提出一种稳健、高效且易于并行的分区优化的混合式SfM方法。首先，利用SIFT算法进行影像匹配，无须GPS/INS等其他辅助信息，仅利用影像间的匹配结果计算得到的影像关联度完成影像分区。然后，提出一种改进的增量式SfM方法实现每个分区内快速重建，以及提出多项标准自动剔除不可靠分区并将这些分区内影像重新划分至其他分区，实现分区的动态调整。最后，提出一种稳健高精度的分区融合算法，实现相机参数、影像姿态和场景三维信息的准确融合。多组不同规模、不同影像类型以及不同场景的典型数据试验结果表明本文方法对不同数据集具有很好的稳健性，在保持高精度的同时能大大提高重建效率，尤其适用于大规模影像数据集。

A hybrid SfM method based on partition optimization

Aiming at solving the problem of sparse 3D reconstruction of large-scale unordered images,this paper proposes a robust,efficient,and easy-to-parallel hybrid SfM method based on partition optimization.Firstly,the SIFT algorithm is used for image matching,and image partitioning is completed using image correlation scores calculated from the matching results without other auxiliary information such as GPS/INS.Secondly,an improved incremental SfM method is applied to achieve rapid reconstruction in each partition,and a number of standards are introduced to automatically eliminate the unreliable partitions and re-divide the images in these partitions into other partitions to achieve dynamic adjustment.Finally,a robust and high-precision partition fusion algorithm is proposed to realize accurate fusion of camera parameters,image posture,and 3D information of the scene.The experimental results of multiple challenging data sets of different scales,different image types,and different scenes show that our proposed hybrid SfM method has good robustness to different data sets and dramatically improves efficiency while maintaining high precision,which especially suitable for large-scale image sets.