顾及多尺度监督的点云语义分割

Point cloud semantic segmentation considering multi-scale supervision

针对复杂场景点云分割精度不高、神经网络隐藏单元缺乏直接监督,难以提取语义明确的点云特征等问题,提出了一种将多尺度监督和SCF-Net相结合的点云语义分割网络。首先构建了一个类别信息生成模块,记录编码器中隐藏单元感受野内的类别,用于解码器中辅助分类器的监督学习。其次将解码阶段的点云类别预测任务分解成一系列点云感受野类别预测任务,通过对解码器中每一层添加辅助分类器,预测当前阶段点云感受野类别,编码阶段生成的类别信息作为标签监督网络学习。模型从粗到细地推理点云感受野类别,最终预测得到点云语义标签。实验结果表明,该方法能够有效提取点云关键信息,提高语义分割精度。

In this paper,a point cloud semantic segmentation network combining multi-scale supervision and SCF-Net is proposed to address the problems of low segmentation accuracy of point cloud in complex scene,the lack of direct supervision in neural network hidden units,and the difficulty in extracting specific point cloud features.A category information generation module is first constructed to record the receptive field categories of hidden unit in the encoder,which is used for the supervised learning of auxiliary classifiers in the decoder.Secondly,the point cloud category prediction task in the decoding stage is decomposed into a series of point cloud receptive field category prediction tasks.By adding auxiliary classifiers to each layer of the decoder,the point cloud receptive field category of the current stage is predicted and the category information generated in the coding stage is used as the label to supervise network learning.The model infers point cloud receptive field categories from coarse to fine,and finally predicts point cloud semantic labels.The experimental results show that the method can effectively extract key information of point cloud and improve the accuracy of semantic segmentation.