基于树形结构编码器的点云补全算法

Point Cloud Completion Algorithm Based on Tree Structure Encoder

点云补全是目前计算机三维视觉领域较为重要的一个方向,目前的深度学习算法采用的是编码器–解码器结构,常用的编码器难以提取出精细的局部特征。本文基于生成对抗网络,提出一种带有树形结构编码器的点云补全算法。树状卷积结构可以提取更为精细的点云特征向量,并提高算法的计算效率。最后利用特征金字塔模型来生成点云的缺失部分。实验结果表明,基于该网络结构补全的点云数据具有有效性,并且补全精度相对于PCN算法有一定精度提高。

Point cloud completion is an important direction in the field of computer 3D vision. The current deep learning algorithm uses an encoder-decoder structure. It is difficult for commonly used encoders to extract fine local features. This paper proposes a method based on generative confrontation network, a neural network structure with a tree structure encoder to automatically repair the shape of the 3D point cloud, and the point cloud is better extracted by tree convolution to extract the feature vector of the point cloud and higher computational efficiency, and finally it uses the feature pyramid model to generate the missing parts of the point cloud. The experimental results show that the point cloud data completed based on the network structure is effective, and the completion accuracy is improved to a certain extent compared with the PCN algorithm.