GA-NET:用于点云语义分割的全局注意力网络
收藏资源简介:
如何从3D点云中学习长距离依赖关系是3D点云分析中的一个挑战性问题。为了解决这个问题,我们提出了一个名为GA-Net的全局注意力网络,用于点云语义分割。这包括一个与点无关的全局注意力模块和一个与点有关的全局注意力模块,以获取3D点云的上下文信息。点独立的全局注意力模块简单地为所有3D点共享一个全局注意力图。在点依赖的全局注意力模块中,对于每一个点,我们利用一个新颖的随机交叉注意力块,只使用两个随机采样的子集来学习所有点的上下文信息。此外,我们设计了一个新的点自适应聚合块来替代线性跳过连接,以聚合更多的区分特征。在三个3D公共数据集上的广泛实验结果表明,我们的方法在大多数情况下都超过了最先进的方法。
Learning long-range dependencies from 3D point clouds is a challenging problem in 3D point cloud analysis. To address this issue, we propose a Global Attention Network named GA-Net for point cloud semantic segmentation. It consists of a point-agnostic global attention module and a point-aware global attention module to capture contextual information from 3D point clouds. The point-agnostic global attention module simply shares a global attention map across all 3D points. In the point-aware global attention module, for each point, we employ a novel random cross-attention block that only uses two randomly sampled subsets to learn contextual information of all points. Furthermore, we design a novel point-adaptive aggregation block to replace linear skip connections for aggregating more discriminative features. Extensive experimental results on three public 3D datasets demonstrate that our method outperforms state-of-the-art approaches in most cases.




