结合伪标签生成与噪声标签学习的弱监督点云分割  

Weakly-Supervised Point Cloud Segmentation Combining Pseudo Label Generation and Noise Label Learning

作　　者：邓安 张鹏 陆竹恒 李蔚清[2] 苏智勇[1] Deng An;Zhang Peng;Lu Zhuheng;Li Weiqing;Su Zhiyong(School of Automation,Nanjing University of Science and Technology,Nanjing 210094;School of computer science and engineering,Nanjing University of Technology,Nanjing 210094)

机构地区：[1]南京理工大学自动化学院,南京210094 [2]南京理工大学计算机科学与工程学院,南京210094

出　　处：《计算机辅助设计与图形学学报》2023年第2期273-283,共11页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家重点研发计划(2018YFB1004904);中央高校基本科研业务费(30918012203).

摘　　要：针对当前基于深度学习的点云分割技术对点级别标注训练数据的依赖问题,提出一种基于伪标签生成与噪声标签学习的弱监督点云分割框架.首先,利用点云具有非局部相似性的特点,基于局部-非局部图对点云数据进行建模;其次,基于关系图卷积网络,以半监督的方式为稀疏标注的点云训练集数据生成高质量的伪标签;最后,针对生成的伪标签中存在噪声标签的问题,设计一种利用含噪声标签数据准确训练现有点云分割网络的渐进式噪声鲁棒损失函数.在ShapeNet Part与S3DIS等公开点云分割数据集上的实验结果表明,该框架在推理阶段不增加模型额外运算量的情况下,当使用10%的标签时,在ShapeNet Part数据集上的分割精度与全监督方法相差0.1%;当使用1%的标签时,在S3DIS数据集上的分割精度与全监督方法相差5.2%.Aiming at the problem that current deep learning-based point cloud segmentation methods require a large amount of dense point-level labeled training data,a weakly-supervised point cloud segmentation framework based on pseudo label generation and noisy label learning is proposed.Firstly,based on the non-local similarity of point clouds,we employ local and non-local graphs with multiple types of edges to model the point cloud.Secondly,we introduce a semi-supervised relational graph convolutional network to generate high-quality pseudo labels for incompletely labeled point cloud training data.Finally,to tackle the noisy labels in generated pseudo labels,a progressive noise-robust loss function is proposed to accurately train the point cloud segmentation network on noisy pseudo labeled data.The proposed framework is evalu-ated on public point cloud segmentation datasets,ShapeNet Part and S3DIS.Experiments show that without adding extra computation in inference stage,the proposed method can achieve segmentation accuracy 0.1%lower than fully supervised method in ShapeNet Part dataset with 10%labels,and 5.2%lower than fully supervised method in S3DIS dataset with 1%labels.

关 键 词：点云分割 弱监督学习 噪声标签学习 图卷积网络 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]