三维点云室内场景的语义实例联合分割网络

Semantic instance joint segmentation network for indoor scene of 3D point cloud

针对点云分割网络无法在复杂的室内场景中实现高精度分割的问题,本文设计了一种基于深度学习的语义实例联合分割网络,同时完成三维点云数据的语义分割和实例分割任务,主要包含多任务学习主干网络、特征融合模块和语义实例特征联合模块等。特征融合模块通过跳跃连接融合多个网络层,分别融合2个任务各自不同级别的特征,加强网络对数据中包含的信息的整合,并选取大型室内场景数据集S3DIS和部件分割数据集ShapeNet进行对比实验。实验结果显示,网络在数据集S3DIS的语义分割的总体准确率为86.5%,在数据集ShapeNet的语义分割类别交并比为83.1%,在数据集S3DIS的实例分割的平均精度为60.8%。语义实例特征联合模块通过多任务级的特征联合增加语义和实例的判别特征,提高了点云的语义分割和实例分割的准确率。

For the problem that the point cloud segmentation network can not realize high-precision segmentation in complex indoor scenarios,this paper designs a joint semantic-instance segmentation network based on deep learning,which can simultaneously complete semantic segmentation and instance segmentation of 3D point cloud data.It mainly includes the multi-task learning backbone network,the feature fusion module and the semantic feature joint instance module.The feature fusion module fuses multiple network layers through hop connection,and fuses the features of two tasks at different levels respectively,so as to strengthen the integration of information contained in the data by the network.The dataset of large indoor scene S3DIS and the component segmentation dataset ShapeNet were selected for comparative experiments.The experimental results show that the overall accuracy of semantic segmentation of the network in the data set S3DIS is 86.5%,the intersection ratio of semantic segmentation categories in the data set ShapeNet is 83.1%,and the average accuracy of instance segmentation in the data set S3DIS is 60.8%.The semantic instance feature combination module increases the discriminant features of semantics and instances through multi-task feature combination,and improves the accuracy of semantic segmentation and instance segmentation of point clouds.