Research On Technologies For 3D Point Cloud Attribute Compression Based On MPEG G-PCC

With the development of 3D scanning technology,3D point cloud can accurately describe 3D objects and scenes,and have been widely used in autonomous driving,virtual reality(VR),free-view broadcasting and etc.A 3D point cloud usually consists of geometry information and attribute information:the geometry information is represented by a set of 3D coordinates,and the attribute information usually includes color,reflectance,normal vector,etc.Although the advantages of 3D point cloud are significant,its huge amount of data brings challenges to the limited transmission bandwidth and storage space.Effective compression technology is urgently needed to achieve highly efficient storage and transmission for 3D point cloud.The broad application prospects of 3D point cloud have attracted widespread attention from researchers and practitioners,and related compression technologies emerged in an endless stream,the most representative of which is the Geometry-based Point Cloud Compression(GPCC)standard proposed by the Moving Picture Experts Group(MPEG).In this work,based on the G-PCC coding framework,attribute prediction optimization methods and quantization improvement methods are proposed for attribute prediction and residual quantization,which significantly improves the attribute compression performance of G-PCC.The details are as follows:1.An attribute prediction method based on the geometry information of neighbor points is proposed.This method assumes that the color components of a point can be represented by a linear combination of geometry information.Based on this assumption,the relationship between geometry information and color component can be established according to the neighboring points,and the color prediction value can be obtained according to the geometry information of the current point.Experimental results show that this method can improve the attribute compression performance of G-PCC to some extent.2.An adaptive prediction mode selection method based on rate-distortion optimization is proposed.We use the sum of absolute reconstruction error as distortion for each prediction mode and the exponential Golomb coding method to estimate the number of coded bits of each prediction mode in this method.After that,the prediction mode with the least rate-distortion cost will be selected for attribute prediction.Experimental results show that this method can improve the attribute compression performance of G-PCC.This method has been adopted by the G-PCC standard and written into the reference software.3.A progressive quantization method based on level of detail(LOD)is proposed.Aiming at the error propagation caused by the LOD-based attribute prediction scheme,this method proposes to set different quantization step sizes for different LOD,i.e.,gradually increase the quantization step size with the LOD order.Experimental results show that this method can greatly improve the attribute compression performance of G-PCC.This method has been adopted by the G-PCC and written into the G-PCC codec description document.4.A point-based adaptive quantization method is proposed.This method calculates the quantization weight which can reflect the influence of each point through the prediction relationship between points.By dividing the quantization step size with the quantization weight,the more influential point will get the smaller quantization step size to guarantee higher reconstruction quality.The smaller reconstruction distortion,and the smaller error propagation caused by the attribute prediction.Experimental results show that this method can significantly improve the attribute compression performance of G-PCC.This method has been adopted by the G-PCC standard and written into the reference software.