Application of Normal Vectors and Color Features in Semantic Segmentation of Colored Point Clouds

Point cloud semantic segmentation is one of the fundamental tasks for 3D scene understanding. The solution based on fully convolutional neural network is the mainstream practice of 3D semantic segmentation. Recent works combining transformers have achieved stunning performance. However, most of these methods focused more on the 3D space coordinates, little attention has been drawn to the normal vectors and color features, which play a significant role in LIDAR point cloud registration and 3D rendering tasks. In this paper, we propose several novel modules utilizing these features to improve the semantic segmentation of colored point clouds. To leverage the normal vector features, we propose an efficient normal vector constraint to supervise the results of semantic segmentation through robust geometric relationships between objects. To harness the color information, we design the relative color encoding to improve the position encoding module in transformer architecture. Furthermore, we introduce a color-domain contrastive learning method that performs neighborhood search in the color domain, enhancing feature learning from distant contexts. Extensive experiments show that our model achieves better performance than baseline method.