This paper introduces submanifold sparse convolutional networks (SSCNs) for efficient processing of high-dimensional, sparse input data, particularly 3D point clouds. SSCNs are designed to handle spatially sparse data more efficiently compared to traditional dense convolutional networks. The authors propose a new type of sparse convolution called submanifold sparse convolution (SSC), which maintains the sparsity structure throughout the network. This allows for the training of deep and efficient networks similar to VGG or ResNet architectures. The performance of SSCNs is evaluated on two tasks: semantic segmentation of 3D point clouds and semantic segmentation of scenes using RGB-D images. The results show that SSCNs outperform state-of-the-art methods in terms of accuracy and computational efficiency, demonstrating the effectiveness of the proposed approach.This paper introduces submanifold sparse convolutional networks (SSCNs) for efficient processing of high-dimensional, sparse input data, particularly 3D point clouds. SSCNs are designed to handle spatially sparse data more efficiently compared to traditional dense convolutional networks. The authors propose a new type of sparse convolution called submanifold sparse convolution (SSC), which maintains the sparsity structure throughout the network. This allows for the training of deep and efficient networks similar to VGG or ResNet architectures. The performance of SSCNs is evaluated on two tasks: semantic segmentation of 3D point clouds and semantic segmentation of scenes using RGB-D images. The results show that SSCNs outperform state-of-the-art methods in terms of accuracy and computational efficiency, demonstrating the effectiveness of the proposed approach.