Entropy Modelling with Voxel Grouping and Cross-Group Attention for Point Cloud Geometry Compression

Point clouds are widely used in 3D applications, driving the need for efficient point cloud compression (PCC) technologies. Deep learning-based PCC methods, particularly octree-based approaches, are popular for their lossless and scalable representation capabilities. However, the exponential increase in octree nodes and spatial complexity with increased depth presents challenges for efficient correlation extraction and coding performance. To address these issues, we propose a voxel-based multi-scale coding method with a novel transformer-based architecture, the Cross-Group Point Cloud Transformer (CGFormer). CGFormer uses relative position coding to capture inter-group correlations between uncoded and already-coded groups within the same scale. It employs a dynamic attention mechanism with a coded group buffer (CGB) to progressively enhance feature extraction and probability estimation. Experimental results demonstrate that our method outperforms state-of-the-art deep learning-based approaches and conventional PCC solutions while maintaining competitive complexity.