PointCartesian-Net: enhancing 3D coordinates for semantic segmentation of large-scale point clouds

Collecting accurate outdoor point cloud data depends on complex algorithms and expensive experimental equipment. The requirement of data collecting and the characteristics of point clouds limit the development of semantic segmentation technology in point clouds. Therefore, this paper proposes a neural network model named PointCartesian-Net that uses only 3D coordinates of point cloud data for semantic segmentation. First, to increase the feature information and reduce the loss of geometric information, the 3D coordinates are encoded to establish a connection between neighboring points. Second, a dense connect and residual connect are employed to progressively increase the receptive field for each 3D point, and aggregated multi-level and multi-scale semantic features obtain rich contextual information. Third, inspired by the success of the SENet model in 2D images, a 3D SENet that learns the relation between the characteristic channels is proposed. It allows the PointCartesian-Net to weight the informative features while suppressing less useful ones. The experimental results produce 60.2% Mean Intersection-over-Union and 89.1% overall accuracy on the large-scale benchmark Semantic3D dataset, which shows the feasibility and applicability of the network. (c) 2021 Optical Society of America