基于s跳图距离重建和旋转等变损失的点云补全网络OACHSSCD

Aims:To address the problem that existing point cloud completion methods focus on restoring local geometric details while neglecting overall structural relationships and robustness to rotational transformations(rotational equivariance),this research proposes a point cloud completion network based on s-hop graph distance reconstruction and rotation equivariance loss.Methods:Firstly,a coarse point cloud generation module based on s-hop graph distance reconstruction was constructed to enhance the feature perception of the global point cloud structure;and a hop graph attention embedding mechanism was introduced to highlight semantically similar adjacent features.Secondly,a rotation equivariance loss regularization term was designed to alleviate the model's sensitivity to rotation transformations.Finally,a high-resolution complete point cloud was generated through a refinement module.Results:Experimental results on PCN,the extension of ShapeNet,and the real dataset KITTI showed that the proposed method outperformed existing methods in CD-T,CD-P,F-Score,and consistency metrics.Conclusions:The proposed method can effectively improve point cloud completion performance,generate point clouds with more reasonable structure and stronger geometric consistency.It has good generalization ability.