基于骨架重建的点云补全网络

Point cloud completion network based on skeleton reconstruction

随着计算机三维视觉的快速发展,包含空间几何信息的点云数据广泛应用于机器人、自动驾驶等场景中,然而由于遮挡、角度受限等原因经常会造成几何语义信息的缺失。为了解决这一问题,提出了SRC-Net,首先利用融合动态图卷积网络编码器和折叠网络解码器的骨架重建网络从残缺点云中重建出几何骨架,接着使用自编码器结构建立几何骨架到均匀完整点云的映射。最后在MVP数据集上的补全结果表明,SRC-Net较现有补全网络可以生成更均匀且光滑分布的高质量完整点云,并能达到更为细节的补全效果,为点云深度学习补全提供了一种新的思路和方法,具有一定指导意义。

With the rapid development of computer 3D vision,point cloud data containing spatial geometric information is widely used in robots,autonomous driving and other scenes.However,due to occlusion,angle limitation and other reasons,geometric information is often missing.In order to solve this problem,SRC-Net is proposed.Firstly,a geometric skeleton is reconstructed from the incomplete point cloud using a designed skeleton reconstruction network that fuses a dynamic graph convolutional network encoder and a folding network decoder,and then the auto-encoder structure is used to establish the mapping from the geometric skeleton to the uniform and complete point cloud.Finally,the completion results on the MVP dataset show that SRC-Net can generate high-quality complete point clouds that are more evenly and smoothly distributed than existing completion networks,and can achieve more detailed completion effects.It provides a new idea and method for point cloud deep learning completion,and has certain guiding significance.