摘要
基于几何的点云压缩(geometry-based point cloud compression,G-PCC)可有效降低点云传输对网络带宽和存储的要求,但重建后的点云质量常因点的消失而显著下降.文章提出了一种基于多分支(multi-branch)的G-PCC点云几何后处理方法,通过提取多尺度几何特征,并在每个尺度上使用基于k近邻的最大池化层来聚合几何邻域信息,从而预测体素块的概率,实现更精确的点云重建.在国际运动图像专家组(Moving Picture Experts Group,MPEG)推荐的通用测试条件下,该方法与G-PCC(octree)、G-PCC(trisoup)相比,平均获得91.89%(84.57%)和75.24%(73.51%)的D1(D2)BD-Rate增益;与传统方法LUT相比,平均获得76.78%(70.37%)的D1(D2)BD-Rate增益;与基于深度学习的方法DGPP相比,平均获得23.95%(21.41%)的BD-Rate增益.此外,该方法相较于现有基于学习的方法,复杂度更低,具有更广阔的应用前景.
Geometry-based point cloud compression(G-PCC)effectively reduces the bandwidth and storage requirements for point cloud transmission.However,the quality of the reconstructed point cloud often degrades significantly due to point disappearance.To address this issue,this paper proposed a multi-branch-based geometry post-processing method for G-PCC compressed point cloud.The method extracted multi-scale geometric features and employed a k-nearest neighbors(k NN)based max pooling layer at each scale to aggregate geometric neighborhood information,thereby predicting voxel occupancy probabilities for more accurate point cloud reconstruction.Evaluated under the Moving Picture Experts Group(MPEG)common test conditions,the proposed method achieved averageBD-Rate gains of 91.89%(84.57%)and 75.24%(73.51%)for D1(D2)metrics compared to G-PCC(octree)and G-PCC(trisoup),respectively.It also obtained average BD-Rate gains of 76.78%(70.37%)for D1(D2)over the traditional LUT method and 23.95%(21.41%)over the deep learning-based DGPP method.Furthermore,the proposed approach demonstrated lower computational complexity compared to existing learning-based methods,indicating promising potential for practical applications.
作者
钱虞杰
丁丹丹
QIAN Yujie;DING Dandan(School of Information Science and Technology,Hangzhou Normal University,Hangzhou 311121,China)
出处
《杭州师范大学学报(自然科学版)》
2025年第6期664-672,共9页
Journal of Hangzhou Normal University(Natural Science Edition)
关键词
点云几何压缩
后处理
深度学习
多尺度
K近邻算法
point cloud geometry compression
post-processing
deep learning
multi-scale
k-nearest neighbors
