The paper introduces a novel deep learning framework for semantic segmentation of large-scale point clouds, addressing the challenges posed by the scale and irregular structure of 3D data. The authors propose the use of superpoint graphs (SPGs), which capture the geometric and contextual relationships between object parts in a compact and rich manner. SPGs are derived from a partition of the point cloud into geometrically homogeneous elements, forming an attributed directed graph where nodes represent simple shapes and edges describe their adjacency with rich features. This representation allows for efficient and accurate semantic segmentation of outdoor and indoor LiDAR scans, achieving state-of-the-art performance on the Semantic3D and S3DIS datasets. The framework consists of PointNets for superpoint embedding and graph convolutions for contextual segmentation, with novel efficient versions of Edge-Conditioned Convolutions and input gating in Gated Recurrent Units. The method significantly improves mIoU scores on both datasets, demonstrating its effectiveness in handling large-scale point cloud data.The paper introduces a novel deep learning framework for semantic segmentation of large-scale point clouds, addressing the challenges posed by the scale and irregular structure of 3D data. The authors propose the use of superpoint graphs (SPGs), which capture the geometric and contextual relationships between object parts in a compact and rich manner. SPGs are derived from a partition of the point cloud into geometrically homogeneous elements, forming an attributed directed graph where nodes represent simple shapes and edges describe their adjacency with rich features. This representation allows for efficient and accurate semantic segmentation of outdoor and indoor LiDAR scans, achieving state-of-the-art performance on the Semantic3D and S3DIS datasets. The framework consists of PointNets for superpoint embedding and graph convolutions for contextual segmentation, with novel efficient versions of Edge-Conditioned Convolutions and input gating in Gated Recurrent Units. The method significantly improves mIoU scores on both datasets, demonstrating its effectiveness in handling large-scale point cloud data.