Spherical coordinate transformation-embedded deep network for primitive instance segmentation of point clouds

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.