This paper introduces submanifold sparse convolutional networks (SSCNs) for efficient processing of high-dimensional, sparse input data, particularly for 3D semantic segmentation of point clouds. The proposed method addresses the inefficiency of traditional convolutional networks on sparse data by developing sparse convolutional operations that maintain spatial sparsity throughout the network. The key contribution is the introduction of submanifold sparse convolution (SSC), a novel convolution operator that keeps the same level of sparsity across layers, enabling the construction of efficient and deep networks for 3D point cloud segmentation. SSCNs outperform existing state-of-the-art methods on the ShapeNet part-segmentation challenge, achieving higher accuracy on the test set compared to other approaches. The method is efficient in terms of computational and memory requirements, as it avoids unnecessary computations on inactive sites. The paper also presents experiments on the NYU Depth dataset, demonstrating that SSCNs achieve higher pixel-wise classification accuracy compared to 2D FCNs, while significantly reducing computational costs. The proposed approach uses sparse convolutions that restrict computations to active sites, allowing for efficient processing of sparse data. The networks are designed to handle data that lies on lower-dimensional submanifolds within higher-dimensional spaces, making them suitable for 3D point cloud segmentation. The implementation includes various network architectures, such as FCNs and U-Nets, adapted for sparse convolutions, and demonstrates their effectiveness in both synthetic and real-world datasets. The results show that SSCNs are not only accurate but also computationally efficient, making them a promising solution for 3D semantic segmentation tasks.This paper introduces submanifold sparse convolutional networks (SSCNs) for efficient processing of high-dimensional, sparse input data, particularly for 3D semantic segmentation of point clouds. The proposed method addresses the inefficiency of traditional convolutional networks on sparse data by developing sparse convolutional operations that maintain spatial sparsity throughout the network. The key contribution is the introduction of submanifold sparse convolution (SSC), a novel convolution operator that keeps the same level of sparsity across layers, enabling the construction of efficient and deep networks for 3D point cloud segmentation. SSCNs outperform existing state-of-the-art methods on the ShapeNet part-segmentation challenge, achieving higher accuracy on the test set compared to other approaches. The method is efficient in terms of computational and memory requirements, as it avoids unnecessary computations on inactive sites. The paper also presents experiments on the NYU Depth dataset, demonstrating that SSCNs achieve higher pixel-wise classification accuracy compared to 2D FCNs, while significantly reducing computational costs. The proposed approach uses sparse convolutions that restrict computations to active sites, allowing for efficient processing of sparse data. The networks are designed to handle data that lies on lower-dimensional submanifolds within higher-dimensional spaces, making them suitable for 3D point cloud segmentation. The implementation includes various network architectures, such as FCNs and U-Nets, adapted for sparse convolutions, and demonstrates their effectiveness in both synthetic and real-world datasets. The results show that SSCNs are not only accurate but also computationally efficient, making them a promising solution for 3D semantic segmentation tasks.