面向点云补全的鲁棒图关注网络

A robust graph attention network for point cloud completion

目的:利用图神经网络与关注机制,构建具有鲁棒性的图关注网络模型并用于点云补全任务。方法:首先,用自关注机制构造局部图的邻接矩阵,不仅考虑中心点与邻域点的关系,而且还考虑邻域点之间的内在相关性,从而有效提取点云的局部几何结构信息。其次,利用邻域中邻域点的特征信息,自适应地更新局部中心点的坐标与特征。此时,以每个点新的坐标组成的点云更能准确描述物体的几何结构细节,增强抗噪能力。最后,为了增强特征提取的鲁棒性,利用残差连接分别融合经过多次微调的点特征作为全局特征,以此来生成点云。结果:与其他点云补全方法在多个常用数据集实验相比,本文构建模型具有最优的补全效果。结论:利用具有鲁棒性的图关注网络模型在点云补全任务中具有先进性。

Aims:The graph neural network and self-attention mechanism were constructed as a robust graph attention network model for point cloud completion tasks.Methods:Firstly,the adjacency matrix of the local graph was constructed by the attention mechanism,which considered the relationship between the center and the neighbor points and the intrinsic correlation between the neighbor points to effectively extract the local geometric structure information of the point cloud.Secondly,the coordinates and features of the local center point were adaptively updated by using the feature information of the neighbor points.At this time,the point cloud composed of the new coordinates of each point could more accurately describe the geometric structure details of the object and enhance the anti-noise ability.Finally,residual connections were used to fuse point features that had been fine-tuned several times as global features to enhance the robustness of feature extraction and to utilize the global features to generate point clouds.Results:Compared with other point cloud completion methods on multiple common datasets,the model constructed in this paper had the best completion effect.Conclusions:The robust graph attention network model is advanced in point cloud completion tasks.