• Voxeldnn
• • our contributions
  • file structure
  • environment
  • command
  • ablation study
  • performance
  • Cite from:
  • contributors

Voxeldnn

Voxeldnn, point cloud compression in deeplearning, migrated to tensorlayer
Voxeldnn, is a kind of PC lossless compression method. The source code uses TensorFlow as deeplearning framework, here we transplant to TensorLayer.

our contributions

1.transplant from TensorFlow to tensorlayer
2.benchmark tests on many PCs, including those not tested by author.
3.ablation tests on downsample depth of 3 and 1, as well as restricted box size of 64 and 32.
4.performance comparasion between TensorLayer and TensorFlow.

file structure

root
└── TensorFlow open source code
└── VoxelDNN_v2_tensorlayer/: Tensorlayer source code
└── Voxeldnn explanation.docx: introduction for paper, migration, performance, instructions
└── Lossless Coding of Point Cloud Geometry using a Deep Generative Model.pdf: paper
└── flowchart.vsdx: flow chart about encoding process
└── trainingset: follow the instructions to prepare datasets

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

cd VoxelDNN_v2_tensorlayer

Training:

python -m training.voxel_dnn_training_tensorlayer-2 -inputmodel Model/voxeldnn64_tl/model_27.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/ -outputmodel Model/voxeldnn64_tl_2/

Encode:

python -m voxel_dnn_coder.voxel_dnn_abac_multi_res_sepa_model-tensorlayer_MP -ply 28_airplane_0270.ply -output Output_tl/ -model64 Model/voxeldnn64_tl/model_48.npz -signaling baseline

Decode:

python -m voxel_dnn_coder.voxel_dnn_abac_multi_res_sepa_model_dec-tensorlayer_MP -level 10 -ply 28_airplane_0270.ply -depth 3 -output Output_tl/ -model64 Model/voxeldnn64_tl/model_48.npz -signaling baseline

For more instructions, please refer to VoxelDNN_v2-main/readme.md

ablation study

As the paper says, VoxelDNN costs too much time in coding. And we think the reason is that for one thing, voxel-based methods would inevitably take more time than other method, as they encodes voxel by voxel. And for another, VoxelDNN needs to traverse all encoding paths to search the best one with lowest bpp, and the traversing also takes much time. Here we make an ablation test to speed up the compression.
Test 1: we restrict the downsample depth to 1 instead of 3, which means the final encoded box size will either be 64 or 32.
Test 2: we set the final encoded box size to be 64.
Test 3: we set the final encoded box size to be 32.
And the results show as below. We can see that compared to the origional result of VoxelDNN_depth3, test 1 gets a little smaller bpov(for less flags) and less time(-7% for most files). Test 2 gets bigger bpov and much less time(-20% for most files). Test 3 gets a little bigger bpov(3%) and the least time(-60%). So test 3 can be seen as the optimized version of VoxelDNN.

performance

We tested on files of different bitwidth on testsets. The results are listed below. Compared with TensorFlow, the bpov of TensorLayer is slighterly better.

Encodedfile	TL_bpov	TF_bpov
sarah_vox10_0023.ply	0.833	0.869
sarah_vox9_0023.ply	0.855	0.879
phil_vox9_0139.ply	0.897	0.919
phil_vox10_0139.ply	0.875	0.909
redandblack_vox10_1550.ply	0.786	0.804
queen_vox10_0200.ply	0.702	0.698
longdress_vox10_1300.ply	0.72	0.738
basketball_player_vox11_00000200.ply	0.667	0.674
loot_vox10_1200.ply	0.699	0.718
dancer_vox11_00000001.ply	0.653	0.661
soldier_vox10_0690.ply	0.718	0.739

Cite from:

@article{nguyen2021lossless,
title={Lossless Coding of Point Cloud Geometry using a Deep Generative Model},
author={Nguyen, Dat Thanh and Quach, Maurice and Valenzise, Giuseppe and Duhamel, Pierre},
journal={arXiv preprint arXiv:2107.00400},
year={2021}
}

contributors

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