Hierarchical Artifact Removal for Encoded Point Clouds With Very Low Bitrate G-PCC Octree

As a pivotal medium for immersive information, point clouds (PCs) inevitably suffer distortion during compression, impacting users’ visual experience. To enhance the visual quality of encoded PCs with very low bitrate G-PCC Octree, a hierarchical artifact removal method is proposed in this paper. In response to the computational complexity in restoring geometry distortion caused by the sparse retention of points during PC compression with very low bitrate G-PCC Octree encoding, octree segmentation is performed on PC to obtain multiple cubes for parallel processing, so as to enhance restoration efficiency. Considering the regularity of the voxelized grid-shaped loss caused by geometry distortion in encoded PCs, a method combining 3D U-Net with attention mechanism and upsampling operations is designed to ext distortion features. Moreover, accounting for the different widths of grid-shaped losses and varying degrees of point reduction induced by different encoding levels, the problem of geometry artifact removal is transformed into a predictive and classificatory challenge of confidence index for PC artifact removal. By partitioning restoration thresholds for different encoding levels, hierarchical artifact removal can be achieved to produce more accurate restoration results. Additionally, as color distortion partly originates from missing points caused by geometry distortion, a strategy involving color information mapping, interpolation and smoothing is presented to efficiently utilize the existing color information to achieve rapid color restoration. Experimental results on PCs from four datasets demonstrate that the proposed method shows advanced performance in both geometry and color artifact removal of distorted PCs with very low bitrate G-PCC Octree encoding.