This paper introduces a novel neural network module called EdgeConv for learning on point clouds. The proposed method dynamically constructs a graph in each layer of the network to capture local geometric structures while maintaining permutation invariance. EdgeConv generates edge features that describe the relationships between points, allowing the model to group points both in Euclidean and semantic space. The method is designed to be compatible with existing architectures and has several appealing properties, including the ability to learn global shape properties through stacking and capturing semantic characteristics over long distances in the original embedding. The authors demonstrate the performance of their model on standard benchmarks including ModelNet40, ShapeNetPart, and S3DIS. The model achieves state-of-the-art performance on these datasets, particularly in classification and segmentation tasks. The key contributions of the work include the introduction of EdgeConv as a novel operation for learning from point clouds, the ability to dynamically update a graph of relationships between points, and the integration of EdgeConv into multiple existing pipelines for point cloud processing. The paper also compares the proposed method with existing approaches, including PointNet and graph CNNs, and shows that EdgeConv outperforms these methods in terms of performance and efficiency. The model is evaluated on various tasks, including classification, part segmentation, and semantic segmentation, and is shown to be robust to partial data and varying point cloud densities. The results demonstrate that the model can effectively capture semantic information from point clouds, even when the points are far apart in the original input space. The method is also shown to be effective in indoor scene segmentation and has potential applications beyond 3D geometry, such as in document retrieval and image processing.This paper introduces a novel neural network module called EdgeConv for learning on point clouds. The proposed method dynamically constructs a graph in each layer of the network to capture local geometric structures while maintaining permutation invariance. EdgeConv generates edge features that describe the relationships between points, allowing the model to group points both in Euclidean and semantic space. The method is designed to be compatible with existing architectures and has several appealing properties, including the ability to learn global shape properties through stacking and capturing semantic characteristics over long distances in the original embedding. The authors demonstrate the performance of their model on standard benchmarks including ModelNet40, ShapeNetPart, and S3DIS. The model achieves state-of-the-art performance on these datasets, particularly in classification and segmentation tasks. The key contributions of the work include the introduction of EdgeConv as a novel operation for learning from point clouds, the ability to dynamically update a graph of relationships between points, and the integration of EdgeConv into multiple existing pipelines for point cloud processing. The paper also compares the proposed method with existing approaches, including PointNet and graph CNNs, and shows that EdgeConv outperforms these methods in terms of performance and efficiency. The model is evaluated on various tasks, including classification, part segmentation, and semantic segmentation, and is shown to be robust to partial data and varying point cloud densities. The results demonstrate that the model can effectively capture semantic information from point clouds, even when the points are far apart in the original input space. The method is also shown to be effective in indoor scene segmentation and has potential applications beyond 3D geometry, such as in document retrieval and image processing.