基于超点图网络的三维点云室内场景分割模型

针对点云数据集样本不均衡及PointNet网络无法充分利用点云邻域信息的问题，提出一种三维点云场景分割模型。根据几何信息将原始点云块同质分割为超点，利用小型PointNet网络将点云原始特征映射到高维空间中，并挖掘场景中深层语义信息。在此基础上，构建自归一化属性门控单元优化点云上下文语义分割效果，采用二维图像领域中的Focal Loss损失函数实现点云场景分割。实验结果表明，该模型在S3DIS数据集上的平均交并比、总体精度、平均精度分别达到63.8%、86.4%、74.3%，较SPG模型分别提升1.7、0.9、1.3个百分点。

Indoor Scene Segmentation Model Using Three-Dimensional Point Cloud Based on Super Point Graph Network

The samples of the existing current point cloud datasets are unbalanced, and PointNet can not effectively utilize the neighborhood information of the point cloud.To address the problem, a three-dimensional scene segmentation method using point cloud is proposed based on super point graph network. Based on geometric information, the original point cloud is divided into several blocks homogeneously.Utilizing small-sized PointNet, the original features of point cloud are mapped to high-dimensional space to mine the deep fine-grained semantic information in the scene.By constructing a kind of self-normalization gate recurrent units, the contextual semantic segmentation performance of the point cloud is improved, and the Focal Loss function in the two-dimensional image field is used to segment the scenes of the point cloud.The experimental results on the S3DIS dataset show that the proposed model exhibits a MIOU of 63.8%, OA of 86.4% and mAcc of 74.3%, which are respectively 1.7%, 0.9% and 1.3% higher than the SPG model.The proposed model significantly improves the semantic segmentation performance of point cloud for three-dimensional scenes.