面向三维模型分割的边界感知点云神经网络

为了能够更好地应用深度神经网络学习三维模型的空间特征,获得更好的三维模型分割效果,提出面向三维模型分割的边界感知点云神经网络.首先,采用边界感知的网格点云化方法,将网格分割问题转化成点云标记问题;然后,利用数据切片方法对转化而来的点云数据进行重采样;最后,利用不同大小卷积核的滤波器提取点云数据的空间特征,并将点云标记的结果对应到原网格模型,得到三维模型分割的结果.在ShapeNetCore数据库上的实验结果表明,该方法不仅能够明显地提高分割的准确率,而且具有边界感知的特性,能够有效地避免过分割现象.

Boundary-Aware Point Based Deep Neural Network for Shape Segmentation

In order to learn spatial features of 3D shapes more effectively by deep neural networks,and have a better performance for shape segmentation,a boundary-aware point-based deep neural network is proposed for shape segmentation.At first,meshes are transformed into points via a boundary-aware method,so that shape segmentation can be treated as a point labelling problem.Then the points are resampled by slicing them into several subsets.At last,point-based deep neural network different kernel size filters are proposed to capture the spatial information from point cloud,and shapes are finally segmented through each point and its related mesh labelled.The experiments on ShapeNetCore datasets show that the proposed approach can obviously improve the accuracy of 3D shape segmentation,and has the boundary-aware property so that over-segmentation is evitable.