基于图像的三维预测及其在水利枢纽中的应用

Image-based 3D prediction and its application in water conservancy projects

为了解决传统基于图像的三维重建中鲁棒性较差、信息获取效率低下的问题,使用了卷积神经网络(convolutional neural networks,CNN),将基于区域的掩模卷积网络(region-based convolutional network method,Mask R-CNN)和图卷积(graph convolutional network,GCN)联合实现三维重建,其中Mask R-CNN完成二维感知GCN实现三维形状推断,该方法不需要进行特征提取与匹配以及复杂的几何运算。通过实验验证了该方法的可行性,采用倒角距离(chamfer distance)及法向量距离作为评价指标与基线系统进行了比较,实验显示,倒角距离缩小了0.2~2.238,法向量距离增大了10.11~36.03,体现了优异性。以水利枢纽图作为实例进行三维重建,为稀疏信息及实例图的三维建模提供了新的思路。

In order to solve the problem of poor robustness and low efficiency of information acquisition in traditional image-based 3D reconstruction,a convolutional neural networks(CNN)was used.The region-based convolutional network(Mask R-CNN)and graph convolutional network(GCN)were combined to complete the two-dimensional perception and three-dimensional shape inference to achieve the three-dimensional reconstruction,without the need for feature extraction and matching and complex geometric operations.The feasibility of the method was verified by experiments.The chamfer distance and normal vector distance were used as evaluation indexes and compared with the baseline system.The experiment shows that the chamfer distance is reduced by 0.2-2.238,and the normal vector distance is increased by 10.11-36.03.It provides a new idea for 3D modeling of sparse information and example diagram.