ANFPCGC++: Point Cloud Geometry Coding Using Augmented Normalizing Flows and Transformer-Based Entropy Model

As immersive media gains increasing prominence, point clouds have emerged as a preferred data representation for presenting complex 3D scenes. However, the large size of point cloud data poses challenges in terms of storage and real-time transmission, prompting the need for highly efficient point cloud compression techniques. In response to these challenges, we introduce a novel approach called ANFPCGC++ (Augmented Normalizing Flow-based Point Cloud Geometry Compression) for lossy static point cloud geometry coding. ANFPCGC++ leverages the power of Augmented Normalizing Flow (ANF) in conjunction with sparse convolution to effectively capture and incorporate spatial correlations inherent in point clouds. ANF offers a higher level of expressiveness compared to conventional methods like variational autoencoders (VAE), resulting in more accurate and faithful latent representations. Furthermore, we introduce a Transformer-based entropy model, that combines the hyperprior and context information, enabling a more precise entropy model that supports parallel computation. Extensive experimental results confirm the superior performance of ANFPCGC++. By comparing to the point cloud coding standards G-PCC and V-PCC, our proposed method achieves remarkable bitrate savings of 63.7% and 60.0% in terms of D1-PSNR, respectively. Additionally, when compared to other deep learning-based point cloud geometry compression methods like PCGCv2 and ANFPCGC, our approach demonstrates an average bitrate reduction of 25.6% and 23.6% in terms of D1-PSNR, respectively. The source code is available at https://github.com/ymnn1996/ANFPCGC2.