【摘要】 点云分析一直以来都是一个具有挑战性的问题，主要是因为点云数据的非结构化特性所致。为了解决这个问题，RepSurf基于PointNet++提出了一种多曲面表示局部点云特征的方法。然而，RepSurf中的集合抽象层仅通过一个MLP学习局部特征，这远远不够。为此，引入了两个模块，即倒置残差模块和注意力模块。这两个简单但有效的即插即用模块可以更好地学习局部特征。倒置残差模块通过添加更多的MLP层，丰富了特征提取过程；而注意力模块则包括通道注意力和空间注意力，更加关注关键点特征的学习，使得学习到的特征更具代表性。在公共基准数据集S3DIS上评估了文中的方法，在语义分割任务中mIoU指标达到72.3%，比RepSurf高出2.5%。更多还原

【Abstract】 Due to the unstructured nature of point cloud data, point cloud analysis has always been challengeable. In view of this, RepSurf proposes an Umbrella RepSurf method for representing local point cloud features based on PointNet++. However,the set abstraction layer in RepSurf learns local features by only one MLP(multilayer perceptron), which is far from sufficient.Therefore, two modules, namely the inverted residual module and the attention module, are proposed. These two simple but effective plug-and-play modules can be used to learn the local features better. The inverted residual module enriches the feature extraction process by adding more MLP layers, while the attention module, including channel attention and spatial attention, pays more attention to the learning of key point features, which makes the learned features more representative. The proposed approach is evaluated on the public benchmark dataset S3DIS. It achieves an mIoU of 72.3% in the task of semantic segmentation, which is 2.5% higher than that of RepSurf.更多还原