Exploring the Relation between Geometric and Semantic Information for Semantic Segmentation of 3D Point Cloud

The emergence of 3D point cloud data has significantly advanced computer vision. Semantic segmentation of 3D real scene point clouds stands out as one of the most formidable challenges in this domain. Despite progress, many existing methodologies suffer from issues such as information loss, redundancy in local contexts, and inadequacies in decoding layer fitting. To tackle these obstacles, a receptive field reconstruction (RFR) mechanism is introduced to address information loss and diminish redundancy in local contexts. Secondly, we present a novel bilateral augmentation module, named the bilateral cross-augment (BCA) module, to bolster the relationship between geometric features and contextual data. Thirdly, an attention-based feature propagation (AFP) module is introduced for upsampling, enabling the adaptive restoration of local features from global context. Additionally, a 3D semantic segmentation network is crafted based on the U-Net architecture. Experimental evaluations are conducted on both the S3DIS large-scale real scene indoor dataset and the SemanticKITTI outdoor benchmarks, yielding state-of-the-art mean Intersection over Union (mIoU) scores compared to recent point-based MLP methods. Furthermore, comprehensive ablation experiments are carried out to showcase the superiority of the proposed modules