基于注意机制的多级点云补全算法OA

In response to the problem of restoring blurry details and low utilization of raw input geometry in some point cloud completion algorithms based on point cloud form,a multi-stage point cloud completion algorithm based on attention mechanism is proposed.Firstly,the original point cloud is input and the point feature matrix is output through a two-layer point network.Then,an attention mechanism is used to reevaluate the weight of the point feature matrix,obtain a new point feature matrix,and aggregate the final global feature vector.Next,the global feature vector is input into the fully connected layer and dilation penalty is used to ob-tain a rough output point cloud.Finally,the rough point cloud generated by the fully connected layer is combined and sampled again with the original input point cloud,and a dense point cloud model is obtained through folding the network.The algorithm designed in this article is tested on the ShapeNet dataset,and the experimental results show that the algorithm can effectively complete point cloud details,it has strong detail restoration ability,and good robustness to noise.