基于自适应聚合与深度优化的三维重建算法

针对现有基于多视图的三维重建方法未充分考虑像素点在其余视图的可见性，从而导致重建完整度不足，且在弱纹理和遮挡区域重建困难等问题，提出了一种应用于高分辨率的三维重建网络。首先提出了一种引入可见性感知的自适应成本聚合方法用于成本量的聚合，通过网络获取视图中像素点的可见性，可以提高遮挡区域重建完整性；采用基于方差预测每像素视差范围，构建空间变化的深度假设面用于分阶段重建，在最后一阶段提出了基于卷积空间传播网络的深度图优化模块，以获得优化的深度图；最后采用改进深度图融合算法，结合所有视图的像素点与3D点的重投影误差进行一致性检查，得到密集点云。在DTU 数据集上与其他方法的定量定性比较结果表明，提出的方法可以重建出细节上表现更好的场景。

Adaptive aggregation and depth refinement multi-view stereo network

In view of the problems that the existing 3D reconstruction methods based on multi views do not fully consider the visibility of pixels in other views, which leads to insufficient reconstruction integrity and difficult reconstruction in weak texture and occlusion areas, this paper proposed an efficient cascaded 3D reconstruction network. Firstly, this paper proposed an adaptive cost aggregation method with visibility awareness for cost aggregation. Then this paper obtained the visibility of the pixels in the view through the network, which could improve the integrity of the occlusion area reconstruction. It predicted the disparity range of each pixel based on variance, and the spatially varying depth hypothesis surface was constructed for staged reconstruction. In the last stage, this paper proposed a depth map optimization module based on convolution spatial propagation network to obtain the optimized depth map. Finally, this paper used an improved depth map fusion algorithm to check the consistency of the re projection errors of the pixels and 3D points in all views, and obtained a dense point cloud. The results of quantitative and qualitative comparison with other methods on DTU data sets show that this method can reconstruct scenes with better performance in detail.