MSVoxelDNN_yehua

Point Cloud Compression, context model, Deep Generative Models, G-PCC, base on VoxelDNN
MSVoxelDNN, is a kind of point cloud lossless compression method, optimized on the basis of VoxelDNN.

our contributions

1.transplant from pytorch to tensorlayer
2.model is not provided by author, so we train MSVoxelDNN’s 24 model and VoxelDNN’s 2 models
3.test with testsets, on pytorch and tensorlayer
4.benchmark and ablation test on downsample depth of 3 and 1

file structure

root
└── MSVoxelDNN-master: tensorlayer code
└── MSVoxelDNN pytorch source code
└── MSVoxelDNN-master/Model: model files on pytorch and tensorlayer
└── MSVoxelDNN explanation.docx: introduction for paper, migration, performance, instructions
└── Multiscale deep context modeling for lossless point cloud geometry compression.pdf: origional paper
└── datasets: see the dataset.zip in dataset part
└── MSVoxelDNN performance.xlsx: detailed performance list

environment

Tensorlayer:

• Ubuntu 18.04
• cuda 10.1.243
• python 3.7.6
• tensorflow-gpu 2.3.1
• tensorlayer3 1.2.0
• refer to requirements.txt

command

1.tensorlayer

VoxelDNN training：

python -m training.voxel_dnn_training_tensorlayer -inputmodel /userhome/MSVoxelDNN/MSVoxelDNN-master/Model/voxeldnn32_tl_test/model_1.npz -dataset /userhome/VoxelDNN/datasets/8iVFBv2/10bitdepth_2_oct4/ -dataset /userhome/VoxelDNN/datasets/CAT1/10bitdepth_2_oct4/ -dataset /userhome/VoxelDNN/datasets/MVUB/10bitdepth_2_oct4/ -dataset /userhome/VoxelDNN/datasets/ModelNet40_200_pc512_oct3/

encode：

python3 -m ms_voxel_dnn_coder.ms_voxel_dnn_encoder_tensorlayer -ply /userhome/PCGCv1/pytorch_eval/28_airplane_0270.ply

ablation study

As the paper shows in table II, the bpov of VoxelDNN is smaller than that of MSVoxelDNN, so in order to improve the performance, we should reduce the downsample depth to 1, and rely more on VoxelDNN coding. We make an ablation test on pytorch version to compare the encoding time and bpov between downsample depth 1 and 3.
The results shows below. We can see from the results that compared with depth=3, depth=1 achieves less encoding time (avg: 2568.8s VS 6777.1s) and smaller bpov (avg: 3.039 VS 4.037)!

performance

• test on testsets, to compare the performance of TensorLayer and Pytorch. From the results below, we can see that the bpov of TL and PT are close, but the encoding time of PT is often less than TL, maybe because of the fast operation of 3D convolution in PT.

Encodedfile	TL_EncTime	TL_bpov	PT_EncTime	PT_bpov
sarah_vox10_0023.ply	625.301	0.871	646.468	0.871
sarah_vox9_0023.ply	156.309	0.854	173.102	0.855
phil_vox9_0139.ply	176.158	0.923	161.953	0.923
phil_vox10_0139.ply	728.739	0.935	710.697	0.935
redandblack_vox10_1550.ply	411.62	0.784	369.163	0.786
queen_vox10_0200.ply	460.982	0.714	359.852	0.713
longdress_vox10_1300.ply	466.588	0.701	382.268	0.703
basketball_player_vox11_00000200.ply	1676.731	0.607	1385.684	0.609
loot_vox10_1200.ply	456.226	0.674	389.951	0.676
dancer_vox11_00000001.ply	1479.507	0.605	1188.119	0.607
soldier_vox10_0690.ply	589.835	0.703	472.866	0.704

Cite from:

@article{nguyen2021multiscale,
title={Multiscale deep context modeling for lossless point cloud geometry compression},
author={Nguyen, Dat Thanh and Quach, Maurice and Valenzise, Giuseppe and Duhamel, Pierre},
journal={arXiv preprint arXiv:2104.09859},
year={2021}
}

contributors

name: Ye Hua
email: yeh@pcl.ac.cn