This paper proposes a novel deep learning framework for semantic segmentation of large-scale point clouds. The framework introduces a representation called superpoint graphs (SPGs), which capture the structure of 3D point clouds by partitioning them into geometrically homogeneous elements. SPGs are represented as attributed directed graphs, where nodes correspond to simple shapes and edges describe their adjacency relationships. The framework uses graph convolutions to exploit the contextual relationships captured in SPGs for semantic segmentation. The proposed method achieves state-of-the-art results on two large-scale point cloud datasets: Semantic3D and S3DIS. For Semantic3D, the method improves mean per-class intersection over union (mIoU) by 11.9 points on the reduced test set and by 8.8 points on the full test set. For S3DIS, the method improves mIoU by 12.4 points. The framework is able to handle large point clouds by decomposing the segmentation task into three distinct problems of different scales, which can be solved with corresponding complexity. The method involves three main steps: geometric partitioning of the point cloud into simple shapes, superpoint embedding using PointNet, and contextual segmentation using graph convolutions. The geometric partitioning is done using a global energy model that ensures the resulting superpoints are semantically homogeneous. Superpoints are then embedded using PointNet, and graph convolutions are used to propagate contextual information for semantic segmentation. The framework also introduces a novel version of Edge-Conditioned Convolutions and a new form of input gating in Gated Recurrent Units to improve the efficiency and effectiveness of the graph convolutions. The method is evaluated on two large-scale point cloud datasets and shows significant improvements in semantic segmentation performance compared to existing methods. The framework is able to handle large point clouds by leveraging the SPG representation and graph convolutions to capture contextual relationships between superpoints.This paper proposes a novel deep learning framework for semantic segmentation of large-scale point clouds. The framework introduces a representation called superpoint graphs (SPGs), which capture the structure of 3D point clouds by partitioning them into geometrically homogeneous elements. SPGs are represented as attributed directed graphs, where nodes correspond to simple shapes and edges describe their adjacency relationships. The framework uses graph convolutions to exploit the contextual relationships captured in SPGs for semantic segmentation. The proposed method achieves state-of-the-art results on two large-scale point cloud datasets: Semantic3D and S3DIS. For Semantic3D, the method improves mean per-class intersection over union (mIoU) by 11.9 points on the reduced test set and by 8.8 points on the full test set. For S3DIS, the method improves mIoU by 12.4 points. The framework is able to handle large point clouds by decomposing the segmentation task into three distinct problems of different scales, which can be solved with corresponding complexity. The method involves three main steps: geometric partitioning of the point cloud into simple shapes, superpoint embedding using PointNet, and contextual segmentation using graph convolutions. The geometric partitioning is done using a global energy model that ensures the resulting superpoints are semantically homogeneous. Superpoints are then embedded using PointNet, and graph convolutions are used to propagate contextual information for semantic segmentation. The framework also introduces a novel version of Edge-Conditioned Convolutions and a new form of input gating in Gated Recurrent Units to improve the efficiency and effectiveness of the graph convolutions. The method is evaluated on two large-scale point cloud datasets and shows significant improvements in semantic segmentation performance compared to existing methods. The framework is able to handle large point clouds by leveraging the SPG representation and graph convolutions to capture contextual relationships between superpoints.