Improved Deep Point Cloud Geometry Compression

Authors: Maurice Quach, Giuseppe Valenzise and Frederic Dufaux
Affiliation: Université Paris-Saclay, CNRS, CentraleSupélec, Laboratoire des signaux et systèmes, 91190 Gif-sur-Yvette, France
Funding: ANR ReVeRy national fund (REVERY ANR-17-CE23-0020)
Links: [Paper]

Please feel free to contact me at maurice.quach /at/ l2s.centralesupelec.fr for discussions or questions.

Experimental data

We provide the full experimental data used for the paper. This includes:

In models, all trained models
In results/data.csv, bitrates and objective metric values for all models, point clouds and metrics (D1/D2)
Compressed and decompressed point clouds for all models (c1 to c6, G-PCC trisoup, G-PCC octree)

Prerequisites

Python 3.6.9
Tensorflow 1.15.0 with CUDA 10.0.130 and cuDNN 7.4.2
tensorflow-compression 1.3
MPEG G-PCC codec mpeg-pcc-tmc13: necessary only to compare results with G-PCC, to obtain more recent versions you need to register on the MPEG Gitlab and request the permissions for MPEG/PCC
MPEG metric software v0.12.3 mpeg-pcc-dmetric: available on the MPEG Gitlab, you need to register and request the permissions for MPEG/PCC
MPEG PCC dataset: refer to Common Test Conditions (CTCs) to download the full dataset, you can also get some point clouds from JPEG Pleno.
packages in requirements.txt

Note 1: using a Linux distribution such as Ubuntu is highly recommended
Note 2: CTCs can be found at wg11.sc29.org in All Meetings > Latest Meeting > Output documents "Common test conditions for point cloud compression". For example, "Common test conditions for PCC", in ISO/IEC JTC1/SC29/WG11 MPEG output document N19324 is in the Alpbach meeting 130.

Getting started

Configuration

Adapt the configurations in ev_experiment.yml to your particular setup:

MPEG_TMC13_DIR: G-PCC folder (mpeg-pcc-tmc13)
PCERROR: mpeg-pcc-dmetric folder
MPEG_DATASET_DIR: MPEG PCC dataset folder
TRAIN_DATASET_PATH: Path to training dataset
TRAIN_RESOLUTION: Resolution of the training dataset
EXPERIMENT_DIR: Experiment folder, all results are saved in this folder, it needs to be created manually

Datasets

The dataset is available in the Git repository as ModelNet40_200_pc512_oct3_4k.zip. If you want to reproduce it, steps are provided below.

First, download the ModelNet40 manually aligned dataset: http://modelnet.cs.princeton.edu.
Then, we generate the training dataset specified in our paper (block size 64) with the following commands:

python ds_select_largest.py ~/data/datasets/ModelNet40 ~/data/datasets/pcc_geo_cnn_v2/ModelNet40_200 200
python ds_mesh_to_pc.py ~/data/datasets/pcc_geo_cnn_v2/ModelNet40_200 ~/data/datasets/pcc_geo_cnn_v2/ModelNet40_200_pc512 --vg_size 512
python ds_pc_octree_blocks.py ~/data/datasets/pcc_geo_cnn_v2/ModelNet40_200_pc512 ~/data/datasets/pcc_geo_cnn_v2/ModelNet40_200_pc512_oct3 --vg_size 512 --level 3 
python ds_select_largest.py ~/data/datasets/pcc_geo_cnn_v2/ModelNet40_200_pc512_oct3 ~/data/datasets/pcc_geo_cnn_v2/ModelNet40_200_pc512_oct3_4k 4000

Parameterized experiments

This section contains examples for experiments parameterized using a configuration file. For manual experiments, please go to the next section.

Run G-PCC experiments:

python mp_run.py ev_experiment.yml

Note 1: It is highly recommended to set EXPERIMENT_DIR on an SSD drive. On an HDD, it is recommended to add the --num_parallel 1 option as it can be extremely slow otherwise.

Train all models, run experiments and compare results:

python tr_train_all.py ev_experiment.yml && \
python ev_run_experiment.py ev_experiment.yml --num_parallel 8 && \
python ev_run_compare.py ev_experiment.yml

Note 1: In the provided configuration ev_experiment.yml, a large number of models are trained. This takes about 4 days on an Nvidia GeForce GTX 1080 Ti. Note 2: Adjust the --num_parallel option for ev_run_experiment.py depending on your GPU memory. With 11GB, we've found 8 to leave a fair amount of GPU memory available.

To build the LaTeX tables and gather the figures:

python ut_build_paper.py ev_experiment.yml figs/

To render the point clouds and the visual comparisons:

python ut_run_render.py ev_experiment.yml

To draw the training plots from tfevents data:

python ut_tensorboard_plots.py ev_experiment.yml

Manual experiments

This section contains usage examples for the individual scripts.

Training:

python tr_train.py \
    ~'/data/datasets/pcc_geo_cnn_v2/ModelNet40_200_pc512_oct3_4k/**/*.ply' \
    ~/datassd/experiments/pcc_geo_cnn_v2/models/c4-ws/1.00e-04 \
    --resolution 64 --lmbda 1.00e-04 --alpha 0.75 --gamma 2.0 --batch_size 32 --model_config c3p

Experiment (compress, decompress, remap colors, compute metrics):

python ev_experiment.py \
    --output_dir ~/datassd/experiments/pcc_geo_cnn_v2/longdress_vox10_1300/c4-ws/1.00e-04 \
    --model_dir ~/datassd/experiments/pcc_geo_cnn_v2/models/c4-ws/1.00e-04 \
    --model_config c3p --opt_metrics d1_mse d2_mse --max_deltas inf \
    --pc_name longdress_vox10_1300 \
    --pcerror_path ~/code/MPEG/mpeg-pcc-dmetric/test/pc_error_d \
    --pcerror_cfg_path ~/code/MPEG/mpeg-pcc-tmc13-v10.0/cfg/trisoup-predlift/lossy-geom-lossy-attrs/longdress_vox10_1300/r06/pcerror.cfg \
    --input_pc ~/data/datasets/mpeg_pcc/Static_Objects_and_Scenes/People/longdress_vox10_1300/longdress_vox10_1300.ply \
    --input_norm ~/data/datasets/mpeg_pcc/Static_Objects_and_Scenes/People/longdress_vox10_1300_n/longdress_vox10_1300_n.ply

Compress/decompress (for D1 and D2 optimized point clouds):

python compress_octree.py \
    --input_files ~/data/datasets/mpeg_pcc/Static_Objects_and_Scenes/People/longdress_vox10_1300/longdress_vox10_1300.ply \
    --input_normals ~/data/datasets/mpeg_pcc/Static_Objects_and_Scenes/People/longdress_vox10_1300_n/longdress_vox10_1300_n.ply \
    --output_files ~/datassd/experiments/pcc_geo_cnn_v2/longdress_vox10_1300/c4-ws/1.00e-04/longdress_vox10_1300_d1.ply.bin \
        ~/datassd/experiments/pcc_geo_cnn_v2/longdress_vox10_1300/c4-ws/1.00e-04/longdress_vox10_1300_d2.ply.bin \
    --checkpoint_dir ~/datassd/experiments/pcc_geo_cnn_v2/models/c4-ws/1.00e-04 \
    --opt_metrics d1_mse d2_mse --resolution 1024 --model_config c3p --octree_level 4 \
    --dec_files ~/datassd/experiments/pcc_geo_cnn_v2/longdress_vox10_1300/c4-ws/1.00e-04/longdress_vox10_1300_d1.ply.bin.ply \
        ~/datassd/experiments/pcc_geo_cnn_v2/longdress_vox10_1300/c4-ws/1.00e-04/longdress_vox10_1300_d2.ply.bin.ply

Compress/decompress (for D1 optimized point cloud only):

python compress_octree.py \
    --input_files ~/data/datasets/mpeg_pcc/Static_Objects_and_Scenes/People/longdress_vox10_1300/longdress_vox10_1300.ply \
    --output_files ~/datassd/experiments/pcc_geo_cnn_v2/longdress_vox10_1300/c4-ws/1.00e-04/longdress_vox10_1300_d1.ply.bin \
    --checkpoint_dir ~/datassd/experiments/pcc_geo_cnn_v2/models/c4-ws/1.00e-04 \
    --opt_metrics d1_mse --resolution 1024 --model_config c3p --octree_level 4 \
    --dec_files ~/datassd/experiments/pcc_geo_cnn_v2/longdress_vox10_1300/c4-ws/1.00e-04/longdress_vox10_1300_d1.ply.bin.ply

Transfer colors from original to decompressed point cloud for visualization:

python map_color.py \
    ~/data/datasets/mpeg_pcc/Static_Objects_and_Scenes/People/longdress_vox10_1300/longdress_vox10_1300.ply \
    ~/datassd/experiments/pcc_geo_cnn_v2/longdress_vox10_1300/c4-ws/1.00e-04/longdress_vox10_1300_d1.ply.bin.ply \
    ~/datassd/experiments/pcc_geo_cnn_v2/longdress_vox10_1300/c4-ws/1.00e-04/longdress_vox10_1300_d1.ply.bin.ply.color.ply

Execute mpeg-pcc-dmetric to compute distortion metrics:

~/code/MPEG/mpeg-pcc-dmetric/test/pc_error_d \
    -a ~/data/datasets/mpeg_pcc/Static_Objects_and_Scenes/People/longdress_vox10_1300/longdress_vox10_1300.ply \
    -b ~/datassd/experiments/pcc_geo_cnn_v2/longdress_vox10_1300/c4-ws/1.00e-04/longdress_vox10_1300_d1.ply.bin.ply \
    -n ~/data/datasets/mpeg_pcc/Static_Objects_and_Scenes/People/longdress_vox10_1300_n/longdress_vox10_1300_n.ply \
    --resolution 1023 --dropdups 2 --neighborsProc 1

Overview

├── requirements.txt                            package requirements
└── src
    ├── compress_octree.py                      [Coding] Compress a point cloud
    ├── decompress_octree.py                    [Coding] Decompress a point cloud
    ├── ds_mesh_to_pc.py                        [Dataset] Convert mesh to point cloud
    ├── ds_pc_octree_blocks.py                  [Dataset] Divide a point cloud into octree blocks
    ├── ds_select_largest.py                    [Dataset] Select the N largest files from a folder
    ├── ev_compare.py                           [Eval] Compare results (curves, BD-Rates and BD-PSNRs...)
    ├── ev_experiment.py                        [Eval] Run pipeline for a point cloud (compress, decompress, eval)
    ├── ev_experiment.yml                       [Config] Experimental configuration
    ├── ev_run_compare.py                       [Eval] Compare results for the experimental configuration
    ├── ev_run_experiment.py                    [Eval] Run pipelines for the experimental configuration
    ├── map_color.py                            [Utils] Transfer colors from a point cloud onto another
    ├── model_configs.py                        [Model] Model configurations
    ├── model_opt.py                            [Model] Threshold optimization
    ├── model_syntax.py                         [Model] Bitsream specification
    ├── model_transforms.py                     [Model] Transforms configurations
    ├── model_types.py                          [Model] Model types
    ├── mp_report.py                            [MPEG] Generate JSON report for a G-PCC folder
    ├── mp_run.py                               [MPEG] Run G-PCC for the experimental configuration
    ├── tr_train.py                             [Train] Train a compression model
    ├── tr_train_all.py                         [Train] Train compression models for the experimental configuration
    ├── ut_build_paper.py                       [Utils] Create LaTeX tables and gather figures
    ├── ut_run_render.py                        [Utils] Render point clouds and visual comparisons
    ├── ut_tensorboard_plots.py                 [Utils] Generate plots from tfevents data
    └── utils
        ├── bd.py                               BD-Rate and BD-PSNR computation
        ├── colorbar.py                         Colorbar generation
        ├── experiment.py                       Experimental utilities
        ├── focal_loss.py                       Focal loss
        ├── matplotlib_utils.py                 Matplotlib utilities
        ├── mpeg_parsing.py                     MPEG log files parsing
        ├── o3d.py                              Open3D utilities (rendering)
        ├── octree_coding.py                    Octree coding
        ├── parallel_process.py                 Parallel processing
        ├── patch_gaussian_conditional.py       Patch tensorflow-compression GaussianConditional to get debug tensors
        ├── pc_io.py                            Point Cloud Input/Output
        ├── pc_metric.py                        Point Cloud geometry distortion metrics (D1 and D2)
        ├── pc_to_camera_params.py              Generate camera parameters for a Point Cloud
        └── pc_to_img.py                        Convert a Point Cloud to an image using predefined camera parameters

Citation

@misc{quach2020improved,
    title={Improved Deep Point Cloud Geometry Compression},
    author={Maurice Quach and Giuseppe Valenzise and Frederic Dufaux},
    year={2020},
    eprint={2006.09043},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}

Cheap and reliable Node.js hosting starts at $3/month, and $1/month static HTML hosting

mauriceqch / pcc_geo_cnn_v2

Programming Languages

Labels

Projects that are alternatives of or similar to pcc geo cnn v2