Various approaches using neural networks have been proposed to address multi-view stereopsis, but most of them lack capabilities to handle large textureless regions. Hence, a compelling matching network learning comprehensive information from stereo images is constructed to enforce smoothness constraints globally. Trained over binocular stereo datasets only, we show that the network can directly handle the DTU multi-view stereo dataset. When merging together multiple depth maps obtained using either stereo matching, an additional point consolidation procedure is often needed for removing outliers and better aligning individual patches. A second network that consolidates 3D point clouds through directly projecting individual 3D points based on point distributions in their neighborhoods is proposed. Unlike the matching network, this network is trained on local information and is scalable for handling point clouds of any sizes and is capable of processing selected areas of interest as well. Quantitative evaluation on the DTU dataset demonstrates our two networks together can generate point clouds comparable to existing state-of-the-art approaches.

已经提出了使用神经网络的各种方法来解决多视图立体视，但是大多数方法都缺乏处理大型无纹理区域的能力。因此，构建了一种从立体图像中学习综合信息的引人入胜的匹配网络，以在全球范围内实施平滑度约束。仅通过双目立体数据集进行训练，我们证明该网络可以直接处理DTU多视图立体数据集。当将使用任一立体匹配获得的多个深度图合并在一起时，通常需要额外的点合并过程以去除异常值并更好地对齐各个面片。提出了第二个网络，该网络通过基于各个3D点在附近的点分布直接投影来合并3D点云。与匹配网络不同，该网络在本地信息上进行了培训，并且可扩展用于处理任何大小的点云，并且还能够处理选定的感兴趣区域。对DTU数据集的定量评估表明，我们的两个网络一起可以生成可与现有最新技术相媲美的点云。