SGSANet: spatial group-based semantic segmentation network for point clouds

Currently, there are numerous voxel-based 3D semantic segmentation strategies, whereas point-based direct methods remain relatively unexplored. In this paper, We have introduced a network architecture with a symmetric design to achieve fine-grained point cloud segmentation. This framework integrates the strengths of CNNs, adaptive kernel density estimation, self-attention mechanisms, and Residual Neural Networks to produce detailed 3D semantic segmentation. Specifically, to enhance local feature extraction and density adaptability, we propose a novel LA_Set Abstraction block, inspired by the Set Abstraction (SA) module of PointNet++. This block comprises both the KSA and SGSA modules. Additionally, we incorporate a Spatial Group Set Attention module to better extract local features and an Inverted Residual module to deepen the network, thereby facilitating the learning of more complex features. Lastly, in the feature propagation layer, the features from the downsampling layer are integrated with those from higher layers using skip connections. These combined features are then processed through an SGSA (Spatial and Group Self-Attention) module to derive the semantic information for each point. The experimental results of the benchmark demonstrate that the model proposed by us achieves a 66.7% mean Intersection over Union on the Area 5 of S3dis.