物件点云的语义实例联合分割网络

Semantic and instance joint segmentation network for point clouds of objects

为同时实现三维物体点云的语义和实例分割,设计了语义与实例分割网络(semantic and instance segmentation network, SISN)。SISN由编码器、解码器和特征融合模块等3部分组成,每部分都包含用于有效学习几何多样性特征的动态图卷积结构。利用编码器对输入的物体点云的语义和实例特征进行统一编码,使用解码器分别对语义和实例特征进行解码。通过特征融合模块进行语义和实例特征信息交互,经Softmax分类和MeanShift聚类得到点云的语义和实例分割结果。PartNet数据集和植物点云数据集上的定性和定量试验结果表明,SISN对物件点云的分割效果优于当前主流网络ASIS和JSNet。

A new semantic and instance segmentation network(SISN)is designed to simultaneously implement semantic and instance segmentation of 3D object point clouds.The SISN consists of three components,including an encoder,a decoder and a feature fusion module,each of which contains a dynamic graph convolution structure for efficient learning of geometrically diverse features.The semantic and instance features of the input object point cloud are encoded using an encoder,and the semantic and instance features are decoded using a decoder.The semantic and instance feature information is interacted through the feature fusion module,and the semantic and instance segmentation results of the point cloud are obtained by Softmax classification and MeanShift clustering.The qualitative and quantitative experimental results on PartNet dataset and plant point cloud dataset show that SISN outperforms the current mainstream networks ASIS and JSNet for segmentation of object point clouds.