NeSPCGC: Neighborhood-Scene Joint LiDAR Point Cloud Geometry Compression

To meet the high efficiency compression demand of LiDAR point cloud data, this paper proposes NeSPCGC, a neighborhood-scene joint point cloud geometry compression method based on a multiscale architecture. In contrast to previous voxel- and octree-based models which are limited to local receptive fields, NeSPCGC leverages the inherent characteristics of LiDAR point clouds by jointly capturing local neighborhood correlations and scene-level structural consistency, achieving more efficient and effective compression. Specifically, NeSPCGC introduces an inter-scale occupancy-guided prior generation unit (POG), which transforms occupancy values from the lower scale into informative features that serve as priors for the current scale. Furthermore, to enhance intra-scale prediction, NeSPCGC employs a joint prediction (JOP) unit comprising two complementary branches: Neighborhood Feature Extractor (NFE), which captures local spatial correlations in 3D space, and Scene Feature Extractor (SFE), which perceives global structural information from the projected 2D range image. The features from both branches are then fused to form a comprehensive representation of the point cloud for occupancy probability prediction. Extensive experiments demonstrate that NeSPCGC achieves BD-BR gains of 35.64\% and 34.62\% over MPEG G-PCC on the SemanticKITTI and Ford datasets, respectively, significantly outperforming state-of-the-art learning-based methods such as EHEM and Unicorn, while delivering much faster coding speed.