by Lu Tao, Wang Limin
@inproceedings{lu2021cga,
title={CGA-Net: Category Guided Aggregation for Point Cloud Semantic Segmentation},
author={Lu, Tao and Wang, Limin and Wu, Gangshan},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
pages={11693--11702},
year={2021}
}
This is the official implementation of CGA-Net: Category Guided Aggregation for Point Cloud Semantic Segmentation, which proposes to utilize different aggregation strategies between the same category and different categories. Here we provide the Tensorflow version. The code is very clear and easy to transplant to other frameworks.
This module can be leveraged in any existing point-based segmentation networks. Here we provide an example of how to apply CGA module to CloserLook3D, please refer to examples/CloserLook3D/README.md.
For other backbones, one can try to modify the source code in CGA/cga.py.
Our Tensorflow code is based on CloserLook3D, RandLA-Net, MeteorNet and we benefit a lot from PointNet2, KPConv.