A point cloud segmentation method based on position encoding and Dual-distance Attention
In recent years,convolution and graph operations have been widely used in research to capture feature infor-mation from point clouds,leading to good performance in semantic segmentation tasks.However,these methods have limitations in representing local information of point clouds,and a significant amount of feature information is lost by employing symmetric pooling operations.To address these issues,the DualRes-Net network is proposed.The network incorporates a Position Encoding Module to encode local coordinate features,enabling the network to focus on point cloud position information and obtain better local feature representations.And differences between center and neigh-boring points are combined with attention using a Dual-distance Attention Pooling,enhancing the adaptive aggregation ability of attention pooling for local point cloud information.A De-Differentiation Residual structure is used in each stage of the network to extract deep features of point clouds.Since different input types have significant distribution differences,MLP is applied to each type of feature separately to stabilize model training and improve model perform-ance.Finally,in the semantic segmentation experiments in S3 DIS Area5,the semantic segmentation performance of the proposed method achieves a mIoU of 63.7％,surpassing many existing networks,and demonstrating the effectiveness of the method.
point cloudsemantic segmentationposition encodingde-differentiation residualdual-distance attention pooling
万方数据
维普
