KNN-based feature learning network for semantic segmentation of point cloud data

Abstract

• A three-part network for semantic segmentation of unorganized point clouds is presented. • The k NN algorithm is utilized to extract local features from the point features. • The point, local, and global features are concatenated for semantic segmentation. • The similarity loss and segmentation loss are integrated in the segmentation network. • Experiments on indoor and outdoor datasets show the promotion in segmentation accuracy. Semantic segmentation of sensed point cloud data plays a significant role in scene understanding and reconstruction, robot navigation, etc. This paper presents a k NN-based 3D semantic segmentation network, which is a structural model for directly processing the unorganized point clouds. The network consists of three modules: point feature extraction, local feature extraction, and semantic segmentation. The first module is designed based on the simplified PointNet to extract powerful high-dimensional point features. Local feature extraction module, the key component of the proposed network, utilizes the k NN algorithm to search k -neighbors of each query point to extract the local and global features. Then the final semantic segmentation part concatenates the extracted features to learn and label the input point clouds. Experimental results on the indoor and outdoor datasets show that the proposed work settles the shortcoming of insufficient local feature extraction of existing models and promotes the accuracy of semantic segmentation.