Large-scale point cloud semantic segmentation via local perception and global descriptor vector

Large-scale point cloud semantic segmentation is a critical aspect of environmental information perception, with far-reaching applications in domains such as auto-driving, remote sensing, and virtual reality systems. Contemporary methodologies for semantic segmentation of point clouds typically employ the K-Nearest Neighbors (KNN) algorithm to learn local features. Nevertheless, this approach introduces a concern regarding local perceptual ambiguity, while effectively capturing global features dispersed across a large-scale scene still presents a significant challenge. To address these limitations, we present LACV-Net, a neural architecture specifically tailored for semantic segmentation of large-scale point clouds. Our LACV-Net comprises three primary elements: (1) The Local Adaptive Feature Augmentation (LAFA) module, which adaptively learns the similarity weight between the local neighbor, thereby enhance local information and mitigate local perception ambiguity. (2) The Aggregation Loss Function, which uses similarity weighted (derived from our LAFA module) neighboring features as offsets, guiding convergence towards centroid features, thereby constraining the similarity weight and further mitigate local perception ambiguity. (3) The Comprehensive Vector of Locally Aggregated Descriptors (C-VLAD) module that seamlessly fuses local features across multiple resolution representations to generate a comprehensive global description vector, thereby capturing global context more efficiently. We have evaluated the performance of LACV-Net with state-of-the-art networks on several benchmarks as S3DIS, Toronto3D, and SensatUrban. The results demonstrate the superior efficacy of LACV-Net in accurately segmenting and classifying large-scale 3D point cloud scenes, highlighting its potential to advance environmental information perception.