Curvature graph convolution for nonuniform point cloud masked autoencoders
A nonuniform grouping and mask strategy are proposed based on curvature graph convolu-tion to further optimize the mask autoencoder.First,curvature graph convolution is proposed to avoid the induction bias caused by fixed neighborhoods.Second,a graph pooling layer is introduced after curvature convolution,which is pooled and grouped according to the local features of point clouds.Finally,the mask probability of each group is learned based on the output features of the pooling layer to avoid redundancy.Experiments show that our method can effectively improve the performance of mask autoencoder in downstream tasks.Our pretrained models achieve 93.7%classification accuracy on ModelNet40 and 5.08 completion accuracy on Completion3Dv2,outperforming current mainstream methods.
万方数据
维普
