基于局部特征聚合网络的三维语义分割

Local Feature Aggregation Networks for 3D Semantic Segmentation

激光雷达采集的自动驾驶场景点云数据规模庞大且包含丰富的空间结构信息,一些方法将点云变换到体素化网格等稠密表示形式进行处理,但却忽略了点云变换引起的信息丢失问题,导致分割性能降低。为此,提出了一种基于局部特征聚合网络的三维语义分割方法。其中的局部特征融合模块,聚合中心点的K个最近点的特征,并通过强大的注意力机制,得到增强的点特征,从而弥补丢失的信息,提高网络的分割精度。此外,为了提高小物体的分类精度,提出了3D注意力特征融合块,通过摒弃常规的特征图拼接,使用注意力机制来决定不同层次语义特征的权重,得到更加丰富的语义特征,提高网络的性能。在SemanticKITTI和nuScenes数据集上的大量实验表明了该方法的优越性。

The point cloud data of autonomous driving scenes collected by LiDAR is large in scale and contains rich spatial structural detail information,and some methods transform the point cloud to dense representations such as voxelization grids for processing,but ignore the information loss and occlusion problems caused by the point cloud transformations,which leads to degradation of segmentation performance.For this reason,this paper proposes a local feature aggregation networks for 3D semantic segmentation.The local feature aggregation module therein aggregates the features of the K nearest points of the center point and obtains enhanced point features through a powerful attention mechanism,thus compensating for the lost information and improving the segmentation accuracy of the network.In addition,in order to improve the classification accuracy of small objects,this paper proposes a 3D attention feature fusion block,which obtains richer semantic features and improves the performance of the network by discarding the conventional feature map splicing and using the attention mechanism to decide the weights of different levels of semantic features.Extensive experiments on SemanticKITTI and nuScenes datasets demonstrate the superiority of the method.