基于位置编码和双距离注意的点云分割方法

A point cloud segmentation method based on position encoding and Dual-distance Attention

近年来,卷积和图运算被广泛用于从点云中捕获特征信息的研究中,在语义分割任务中表现出良好的性能。然而,这些方法在表示点云的局部信息方面存在局限性,并且采用对称池化操作而丢失了大量的特征信息。为了解决这些问题,提出DualRes-Net网络。该网络采用位置编码模块对局部坐标特征进行编码,使网络能够专注于点云位置信息,获得更好的局部特征表示。利用双距离注意力池将中心点与邻近点的差异与注意力相结合,增强了注意力对局部点云信息的自适应聚合能力。在网络的每个阶段使用去分化残差结构来提取点云的深层特征。由于不同的输入类型具有显著的分布差异,为了稳定模型训练,提高模型性能,对每种类型的特征分别应用MLP。在S3DIS Area5的语义分割实验中,所提方法的分割性能mIoU达到了63.7%,超过了许多现有的网络,证明了所提方法的有效性。

In recent years,convolution and graph operations have been widely used in research to capture feature information 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 neighboring 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 performance.Finally,in the semantic segmentation experiments in S3DIS 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.