OctAttention: Octree-Based Large-Scale Contexts Model for Point Cloud Compression. AAAI 2022 Paper.

November 7, 2023 · View on GitHub

Branches

There are two branches named obj and lidar that implement Object and LiDAR point cloud coding respectively. They share the same network. Note: the checkpoint file is saved in the corresponding branch separately. The model for LiDAR compression is here.

Requirements

  • python 3.7
  • PyTorch 1.9.0+cu102
  • file/environment.sh to help you build this environment

Download and Prepare Training and Testing Data

  • Download data

    For LiDAR compression

    SemanticKITTI (80G)
    23201/20351 frames in 00-10/11-21 folders for training/testing.

    For Object compression

    MPEG 8iVFBv2 (5.5GB)
    300/300 frames in soldier10 and longdress10 for training.
    300/300 frames in loot10 and redandblack10 for testing.

    MPEG 8iVSLF (100M)
    1/1/1/1 frame in Boxer9/10 and Thaidancer9/10 (quantized from 12bit data) for testing.
    please cite: Maja Krivokuća, Philip A. Chou, and Patrick Savill, “8i Voxelized Surface Light Field (8iVSLF) Dataset,” ISO/IEC JTC1/SC29 WG11 (MPEG) input document m42914, Ljubljana, July 2018.

    JPEG MVUB (8GB)
    318/216/207 frames in andrew10, david10 and sarah10 for training.
    245/245/216/216 frames in Phil9/10 and Ricardo9/10 for testing.
    (Note: We rotated the MVUB data to make it consistent with MPEG 8i. Please set rotation=True in the dataPrepare function when processing MVUB data in training and testing.)

  • Prepare data

Please set oriDir in dataPrepare.py before.

python dataPrepare.py

To prepare train and test data. It will generate *.mat data in the directory Data.

Train

python octAttention.py

You should set the Network parameters expName,DataRootetc. in networkTool.py. This will output checkpoint in expName folder, e.g. Exp/Kitti. (Note: You should run DataFolder.calcdataLenPerFile() in dataset.py for a new dataset, and you can comment it after you get the parameter dataLenPerFile)

Encode and Decode

You may need to run the following command to provide pc_error and tmc13v14_r(release version) execute permission.

chmod +x file/pc_error file/tmc13v14_r 

python encoder.py

This will output binary codes saved in .bin format in Exp(expName)/data, and will generate *.mat data in the directory Data/testPly.

python decoder.py

This will load *.mat data for check and calculate PSNR by pc_error.

Test TMC

We provide the test code for TMC13 v14 (G-PCC) for Object and LiDAR point cloud compression.

python testTMC.py

Citation

If this work is useful for your research, please consider citing :

@article{OctAttention, 
title={OctAttention: Octree-Based Large-Scale Contexts Model for Point Cloud Compression}, volume={36}, 
url={https://ojs.aaai.org/index.php/AAAI/article/view/19942}, DOI={10.1609/aaai.v36i1.19942},
number={1}, journal={Proceedings of the AAAI Conference on Artificial Intelligence}, 
author={Fu, Chunyang and Li, Ge and Song, Rui and Gao, Wei and Liu, Shan}, year={2022}, month={Jun.}, pages={625-633} 
}