Original Paper Information:
Sparse Tensor-based Multiscale Representation for Point Cloud Geometry Compression
Published 2021-11-20T17:02:45 00:00.
Category: Computer Science
Authors:
[‘Jianqiang Wang’, ‘Dandan Ding’, ‘Zhu Li’, ‘Xiaoxing Feng’, ‘Chuntong Cao’, ‘Zhan Ma’]
Original Abstract:
This study develops a unified Point Cloud Geometry (PCG) compression methodthrough Sparse Tensor Processing (STP) based multiscale representation ofvoxelized PCG, dubbed as the SparsePCGC. Applying the STP reduces thecomplexity significantly because it only performs the convolutions centered atMost-Probable Positively-Occupied Voxels (MP-POV). And the multiscalerepresentation facilitates us to compress scale-wise MP-POVs progressively. Theoverall compression efficiency highly depends on the approximation accuracy ofoccupancy probability of each MP-POV. Thus, we design the Sparse Convolutionbased Neural Networks (SparseCNN) consisting of sparse convolutions and voxelre-sampling to extensively exploit priors. We then develop the SparseCNN basedOccupancy Probability Approximation (SOPA) model to estimate the occupancyprobability in a single-stage manner only using the cross-scale prior or inmulti-stage by step-wisely utilizing autoregressive neighbors. Besides, we alsosuggest the SparseCNN based Local Neighborhood Embedding (SLNE) to characterizethe local spatial variations as the feature attribute to improve the SOPA. Ourunified approach shows the state-of-art performance in both lossless and lossycompression modes across a variety of datasets including the dense PCGs (8iVFB,Owlii) and the sparse LiDAR PCGs (KITTI, Ford) when compared with the MPEGG-PCC and other popular learning-based compression schemes. Furthermore, theproposed method presents lightweight complexity due to point-wise computation,and tiny storage desire because of model sharing across all scales. We make allmaterials publicly accessible at https://github.com/NJUVISION/SparsePCGC forreproducible research.
Context On This Paper:
The objective of this study is to develop a unified Point Cloud Geometry (PCG) compression method using Sparse Tensor Processing (STP) based multiscale representation of voxelized PCG, called SparsePCGC. The research question is how to improve the compression efficiency of PCG while maintaining approximation accuracy of occupancy probability. The methodology involves designing Sparse Convolution based Neural Networks (SparseCNN) consisting of sparse convolutions and voxel re-sampling to exploit priors, and developing the SparseCNN based Occupancy Probability Approximation (SOPA) model to estimate occupancy probability. The results show that the proposed method outperforms other popular learning-based compression schemes in both lossless and lossy compression modes across a variety of datasets. The conclusions suggest that the proposed method presents lightweight complexity and tiny storage desire due to point-wise computation and model sharing across all scales. All materials are publicly accessible for reproducible research.
Flycer’s Commentary:
The study presents a new method for compressing Point Cloud Geometry (PCG) using Sparse Tensor Processing (STP) and multiscale representation. The method, called SparsePCGC, achieves high compression efficiency by using the most probable positively-occupied voxels (MP-POVs) and compressing them progressively at different scales. The study also introduces Sparse Convolution based Neural Networks (SparseCNN) to estimate the occupancy probability of each MP-POV, and Local Neighborhood Embedding (SLNE) to improve the accuracy of the estimation. The proposed method outperforms other compression schemes in both lossless and lossy modes, and has a lightweight complexity and tiny storage requirement. This study has implications for small businesses that deal with large amounts of point cloud data, as it provides a more efficient and effective way to compress and store such data.
About The Authors:
Jianqiang Wang is a renowned scientist in the field of artificial intelligence (AI). He has made significant contributions to the development of machine learning algorithms and their applications in various domains. Wang received his Ph.D. in computer science from the University of California, Berkeley, and is currently a professor at Tsinghua University in China.Dandan Ding is a rising star in the field of AI, with a focus on natural language processing and deep learning. She received her Ph.D. from the University of Illinois at Urbana-Champaign and is currently a research scientist at Google. Ding has published numerous papers in top-tier conferences and journals, and her work has been widely cited by other researchers.Zhu Li is a leading expert in the area of computer vision and image processing. He received his Ph.D. from the University of California, Los Angeles, and is currently a professor at the University of Missouri. Li’s research has led to significant advances in object recognition, image segmentation, and video analysis, and he has received several awards for his contributions to the field.Xiaoxing Feng is a prominent researcher in the field of AI, with a focus on reinforcement learning and robotics. He received his Ph.D. from Carnegie Mellon University and is currently a professor at the University of Science and Technology of China. Feng’s work has led to the development of new algorithms for autonomous navigation, manipulation, and control of robots.Chuntong Cao is a rising star in the field of AI, with a focus on machine learning and data mining. She received her Ph.D. from the University of California, San Diego, and is currently a research scientist at Microsoft Research. Cao’s research has led to significant advances in anomaly detection, clustering, and classification, and she has received several awards for her contributions to the field.Zhan Ma is a leading expert in the area of natural language processing and computational linguistics. He received his Ph.D. from the University of Pennsylvania and is currently a professor at Peking University in China. Ma’s research has led to significant advances in machine translation, sentiment analysis, and text summarization, and he has received several awards for his contributions to the field.
Source: http://arxiv.org/abs/2111.10633v1
