DANet: Density-Adaptive Network for Geometry-based Point Cloud Compression Artifacts Removal
Zetao Yang, Wei Gao, Xijing Lu
Abstract
To remove compression artifacts in voxelized point clouds compressed by Geometry-based Point Cloud Compression (G-PCC), we propose the Density-Adaptive Network (DANet). The DANet considers the density characteristics of decompressed point clouds and density variation within network processing, which facilities the removal of artifacts. Regarding density characteristics, we design the corresponding modules, including Sparse-convolution Point Aggregation (SPA), Dense Convolution Block (DCB) and Dilated Hyper-Cross Block (DHCB). The proposed SPA combines point-based feature extraction and sparse convolution, enabling efficient processing of both moderately and intensively compressed point clouds. The proposed DCB and DHCB effectively cater to the density variation during network processing, expanding receptive fields while maintaining an affordable computational cost. Experimental results on representative datasets demonstrate that our proposed method can bring a 91.6% D1 BD-Rate gain over G-PCC. Compared with advanced deep learning-based methods, DANet also achieves 4.6% D1 BD-Rate improvement with a smaller model size.