影像一致性优化的视图选取策略

在三维重建中，网格优化通常用于解决密集点云构建的三角网格含有较多噪声且缺少细节的问题。现有的变分优化方法利用完整的影像数据对初始网格进行影像一致性优化，但在一定程度上忽视了影像信息的冗余以及视图的质量对网格优化的影响。对此，本文提出主视图选取与从视图选取策略，以提升网格优化的效率与质量。首先综合影像梯度幅值与轮廓检测，构建马尔科夫随机场，为每个三角面选取主视图；其次根据相应的观测条件为每个主视图选取从视图；然后计算主、从视图间归一化加权的影像一致性；最后利用梯度下降法最小化表面能量函数，实现网格优化。试验在定性和定量上证实了本文方法的有效性，表明本文方法能恢复更多精细细节，且优化的时间更短、精度更高。

View selection strategy for photo-consistency refinement

In 3D reconstruction, mesh refinement is generally applied to deal with noise and lack of details in triangular mesh built from dense point cloud. The existing variational refinement methods optimize the photo-consistency of the initial mesh by utilizing all the image data, but ignore the redundancy of image information and the impact of view quality on mesh refinement to some extent. In this regard, this paper proposes the strategies of master view selection and slave view selection, so as to improve the efficiency and quality of mesh refinement. Firstly, Markov random field is constructed by combining image gradient magnitude and contour detection to select master view for each triangular facet, and then slave view is selected according to the corresponding observation condition for each master view. Afterwards, we calculate the norm-weighted photo-consistency between master view and slave view, and finally surface energy function is minimized by using gradient decent method to obtain the refined mesh. The experiments show that the proposed method can recover more fine-scale details, meanwhile shorten time and increase accuracy of refinement, which confirms the validity of the proposed method qualitatively and quantitatively.