OctPCGC-Net: Learning Octree-Structured Context Entropy Model for Point Cloud Geometry Compression

Abstract

In Point Cloud Geometry Compression (PCGC), an accurate context entropy model is necessary to reduce spatial redundancy. The octree-based auto-regressive context entropy model has great potential to explore large-scale context dependency. However, over-concentrated attention maps and instability of training process usually occur in large-scale context entropy models. To address these problems, we propose a novel OctPCGC-Net for PCGC based on deep learning framework. Specifically, we introduce a scaled cosine attention method in a large-scale context entropy model to alleviate the problem of over-concentrated attention maps caused by self-attention mechanism, thereby improving the model's prediction accuracy. In order to improve the stability of model training, we further introduce a residual post normalization strategy to alleviate the phenomenon of accumulating activation scores as the network deepens, which makes the activation scores of different layers smoother and more stable. Experimental results show that compared with the state-of-the-art large-scale auto-regressive entropy models, our method saves 6.3%, 8.7%, and 6.3% bitrates in terms of Bjøntegaard Delta Bit Rate (BDBR) on benchmark datasets SemanticKITTI, 8iVFB, and Owlii, respectively. Additionally, our method also achieves higher reconstruction quality (D1 PSNR) and smaller Chamfer distance (CD) under similar bits per point (BPP) on SemanticKITTI dataset.