Learned Point Cloud Geometry Compression

With the increasing demand for volumetric visual data,images cannot well rep-resent the real three-dimensional(3D)world.In contrast,Point Cloud(PC)use 3D point set to directly and efficiently represent 3D objects and scenes,which has been widely used in immersive media as well as 3D sensing for autonomous driving.How-ever,PC has huge data volume and irregular data structure,which limits its storage and transmission efficiency,efficient data compression methods for PC are highly disered.Target the subject of Point Cloud Compression(PCC),we review the background and various PCC technologies.Inspired by the advanced learning-based compression and PC processing methods,we proposed the Learned Point Cloud Geometry Compression(Learned-PCGC)approach,which is novel and achieves the state-of-the-art compres-sion performance.In this article,we first present a contemporary survey of different PCC approaches.We review advanced technologies such as MPEG(Moving Picture Experts Group)PCC standard.Specifically,we introduce octree coding,triangle mesh coding,projection-based coding,as well as the PC attribute compression.We also review the related learning based data compression and point cloud processing methods,such as end-to-end learning based image compression,point cloud neural network,and the emerging learning based PCC methods.Then we introduce the proposed Learned lossy Point Cloud Geometry Compres-sion framework(Learned-PCGC),which is based on the volumetric representation.The overall framework consists of a pre-processing module for point cloud voxelization and partition,a compression network for rate-distortion optimized representation,and a post-processing module for point cloud reconstruction,to systematically represen-t Point Cloud Geometry(PCG)using compact latent features.we conduct sufficient ablation studies to demonstrate the efficiency of the proposed novel tools such as s-caling in pre-processing,deep compression neural network,and adaptive thresholding for classification in post-processing.We also perform a lot of comparisons with other PCGC methods,showing excellent results on dense point clouds.Our method achieves more than 60%BD-Rate gains against popular MPEG Geometry-based PCC(G-PCC).We further propose a more efficient multiscale Learned-PCGC method,which has better compression efficiency and lower complexity.The method is based on sparse tensor representation rather than redundant volumetric data.And the framework is developed on top of a sparse convolution based Auto Encoder(AE)for multiscale re-sampling.For the input PCG data,the encoder network translates it to a downscaled point cloud at the bottleneck layer which possesses both geometry and associated fea-ture attributes.Then the geometry and feature attributes are separately compressed.At the decoder network,we use a binary classification based hierarchical reconstruction to progressively decode the feature attribute and reconstruct the origin PCG information.Using the proposed multiscale sampling and reconstruction with sparse convolution,we can leverage the sparsity nature of the PC for compact and efficient feature representa-tion.Experimental comparison has shown that our method can achieve>30%BD-Rate gains over the previous Learned-PCGC method,and outperforms other traditional PCC methods,achieves state-of-the-art compression efficiency.In addition to the above work about lossy PCGC,we also studied learned lossless geometry compression and color attribute compression.In the future,we will continue to explore more topics such as dynamic PCGC.We have made all materials of our work publicly for research??.