基于对比边界学习的超面片Transformer点云分割网络

Contrastive boundary learning based Transformer network for point cloud segmentation

为了高效地实现大规模室内点云场景语义分割,针对边界点云的特征信息难以区分、场景点云规模过于庞大而导致其难以直接进行分割网络的有效训练等问题,以超面片为数据表征,结合超面片Transformer模块(SPT)和对比边界学习模块(CBL),提出一种基于对比边界学习的超面片Transformer点云分割网络。针对数据集S3DIS进行训练,实验结果表明,该网络在分割精度上比Dgcnn网络高3.9%,在训练速度方面比SPGraph网络快近100倍,针对大规模室内点云场景分割效果尤为突出。

For the issue of semantic segmentation for large-scale indoor point clouds scenes,it is difficult to distinguish the feature information of boundary point clouds and it is challenging to train deep neural networks efficiently due to the vast amount of point clouds data.Taking scene super-patch as data representation,combining the super-patch Transformer(SPT)module and the contrastive boundary learning(CBL)module,a contrastive boundary learning based Transformer network is proposed.The network is trained on public datasets S3DIS.The experimental results show that the overall accuracy of the network is 3.9%higher than the Dgcnn network,and the network training speed is nearly 100 times faster than the SPGraph network.The segmentation effect is outstanding in large-scale indoor point cloud scenes.