NGS_dengy02
Introduction to NGS
The NGS comprises of three consecutive steps: first, constructing a local graph of each point using its K nearest neighbors based on the Euclidean distance metric; second, for each local graph, aggregating neighbor weights using point-wise dynamic filter to the graph center point as its feature attribute, by which embedding local structural/geometric variations as the latent features; finally, devising attention-based sampling to all points that having neighboring structures aggregated, to select a subset of points for compact and precise representation of input point cloud.
source code is here: https://github.com/linyaog/point_based_pcgc
Point cloud geometry (PCG), local neighborhood graph, dynamic filter, attention-based sampling, point-wise convolution
Test indicators
- Using the shapenetcorev2 dataset used in the paper, four indicators of its 55 categories are tested: Test time (each point cloud), bpp, D1-PSNR, D2-PSNR.
- Except for the two point cloud categories car and rifle in the paper, the indicators of other categories are selected as the best test values.
Contributions
-
According to different hyperparameters, 12 optimal models were trained (referred to the readpoints folder in detail).
-
According to the optimal model, the Test Times (Each Point Cloud), bpp, D1-PSNR, D2-PSNR of all 55 categories of Shapenetcorev2 DataSet are tested.
performance
The test results are shown in the following table:
Note that my hardware configuration is different from Paper:
The GPUs of the paper: Nvidia GeForce GTX 1080 Ti.
The CPU of the paper: an Intel Xeon CPU E5-2683 v4.
The GPUs of my test: TESLA T4.
The CPU of my test : Intel(R) Xeon(R) Gold 6248 CPU.
Citation
@INPROCEEDINGS{9506631, author={Gao, Linyao and Fan, Tingyu and Wan, Jianqiang and Xu, Yiling and Sun, Jun and Ma, Zhan}, booktitle={2021 IEEE International Conference on Image Processing (ICIP)}, title={Point Cloud Geometry Compression Via Neural Graph Sampling}, year={2021}, volume={}, number={}, pages={3373-3377}, doi={10.1109/ICIP42928.2021.9506631}}
Contributor
name: Deng Yu
email: dengy02@pcl.ac.cn