In this research, a primitive prediction network embedding Spherical Coordinate Transformation (named SCT-Net), which is a simple and end-to-end deep neural network, is proposed for primitive instance segmentation of point clouds. The key point of SCT-Net is to excavate the relationship between local neighborhood points. First, in order to enhance the compacted expression of local feature, a spherical coordinate transformation is embedded to a deep network. Second, the embedded network is constructed to predict the point grouping proposals and classify the primitives corresponding to each proposal, which can segment primitive instance directly. Third, the feature relationship between each two points is revealed by the constructed relation matrix. The designed loss function not only encourages the embedded network to describe local surface properties, but also produces a grouping strategy accurately for each point. Experiments show that the proposed SCT-Net achieves the state-of-the-art performance than representative methods. At the same time, the capability of spherical coordinate transformation has been demonstrated to improve primitive instance segmentation.

在这项研究中，提出了一种嵌入球坐标变换的原始预测网络（称为 SCT-Net），它是一种简单的端到端深度神经网络，用于点云的原始实例分割。SCT-Net的关键在于挖掘局部邻域点之间的关系。首先，为了增强局部特征的压缩表达，将球坐标变换嵌入到深度网络中。其次，构建嵌入式网络来预测点分组提议，并对每个提议对应的图元进行分类，可以直接分割图元实例。第三，通过构建的关系矩阵揭示每两点之间的特征关系。设计的损失函数不仅鼓励嵌入式网络描述局部表面属性，而且还为每个点准确地生成分组策略。实验表明，所提出的 SCT-Net 实现了比代表性方法最先进的性能。同时，球坐标变换的能力已被证明可以改善原始实例分割。