Multiscale Point Cloud Geometry Compression

​ We apply an end-to-end learning framework to compress the 3D point cloud geometry (PCG) efficiently. Leveraging the sparsity nature of point cloud, we introduce the multiscale structure to represent native PCG compactly, offering the hierarchical reconstruction capability via progressive learnt re-sampling. Under this framework, we devise the sparse convolution-based autoencoder for feature analysis and aggregation. At the bottleneck layer, geometric occupancy information is losslessly encoded with a very small percentage of bits consumption, and corresponding feature attributes are lossy compressed.



We recommend you to follow https://github.com/NVIDIA/MinkowskiEngine to setup the environment for sparse convolution.



Please download the pretrained models and install tmc3 mentioned above first.

sudo chmod 777 tmc3 pc_error_d
python coder.py --filedir='longdress_vox10_1300.ply' --ckptdir='ckpts/r3_0.10bpp.pth' --scaling_factor=1.0 --rho=1.0 --res=1024
python test.py --filedir='longdress_vox10_1300.ply' --scaling_factor=1.0 --rho=1.0 --res=1024
python test.py --filedir='dancer_vox11_00000001.ply'--scaling_factor=1.0 --rho=1.0 --res=2048
python test.py --filedir='Staue_Klimt_vox12.ply' --scaling_factor=0.375 --rho=4.0 --res=4096
python test.py --filedir='House_without_roof_00057_vox12.ply' --scaling_factor=0.375 --rho=1.0 --res=4096

The testing rusults of 8iVFB can be found in ./results


 python train.py --dataset='training_dataset_rootdir'


These files are provided by Nanjing University Vision Lab. And thanks for the help from Prof. Dandan Ding from Hangzhou Normal University and Prof. Zhu Li from University of Missouri at Kansas. Please contact us (mazhan@nju.edu.cn and wangjq@smail.nju.edu.cn) if you have any questions.