GA-NET: Global Attention Network for Point Cloud Semantic Segmentation

We propose a global attention network, called GA-Net, to obtain global information of point clouds in an efficient way. GA-Net consists of a point-independent global attention module, and a point-dependent global attention module.

[PDF]

Overview

This repository is the author's re-implementation of GA-Net. Extensive experiments on three point cloud semantic segmentation benchmarks demonstrate that GA-Net outperforms state-of-the-art methods in most cases.

Further information please contact Shuang Deng and Qiulei Dong.

Citation

Please cite this paper if you want to use it in your work:

@article{2021ganet,
title={GA-NET: Global Attention Network for Point Cloud Semantic Segmentation}, 
author={Deng, Shuang and Dong, Qiulei},
journal={IEEE Signal Processing Letters (SPL)}, 
volume={28},
pages={1300-1304}, 
year={2021},
doi={10.1109/LSP.2021.3082851}
}

Setup

Setup python environment:

conda create -n ganet python=3.6
source activate ganet
pip install -r helper_requirements.txt
sh compile_op.sh

Semantic3D

Download and extract the Semantic3D dataset:

sh utils/download_semantic3d.sh

Prepare the Semantic3D dataset:

python utils/data_prepare_semantic3d.py

Train:

python main_Semantic3D.py --gpu $your_gpu_id --mode 'train'

Evaluation:

python main_Semantic3D.py --gpu $your_gpu_id --mode 'test'

The trained model is stored in the folder result/ganet/Log_2020-10-09_Semantic3D_1.

S3DIS

Download the S3DIS dataset from here (4.09GB). Uncompress the folder and move it to /your_data_folder/S3DIS.

Prepare the S3DIS dataset:

python utils/data_prepare_s3dis.py

Train:

python main_S3DIS.py --model 'GANet' --test_area 5 --gpu $your_gpu_id --mode 'train'

Test:

python main_S3DIS.py --model 'GANet' --test_area 5 --gpu $your_gpu_id --mode 'test'

Calculate the final mean IoU results:

python utils/area_5_cv.py

The trained model is stored in the folder result/ganet/Log_2020-10-19_S3DIS_Area_5.

ScanNet

Download the ScanNet dataset from here (1.72GB). Uncompress the folder and move it to /your_data_folder/scannet.

Prepare the ScanNet dataset:

python utils/data_prepare_scannet.py

Train:

python main_ScanNet.py --model 'GANet' --gpu $your_gpu_id --mode 'train'

Test:

python main_ScanNet.py --model 'GANet' --gpu $your_gpu_id --mode 'test'

The trained model is stored in the folder result/ganet/Log_2020-10-09_ScanNet_1.

Acknowledgement

The structure of this codebase is borrowed from RandLA-Net.