attentions_in_3D_detection
This repository is for the following paper: "Investigating Attention Mechanism in 3D Point Cloud Object Detection"
Shi Qiu*, Yunfan Wu*, Saeed Anwar, Chongyi Li
Paper and citation
The paper can be downloaded from here (arXiv).
If you find our paper/codes/results are useful, please cite:
@article{qiu2021investigating,
title={Investigating Attention Mechanism in 3D Point Cloud Object Detection},
author={Qiu, Shi and Wu, Yunfan and Anwar, Saeed and Li, Chongyi},
journal={arXiv preprint arXiv:2108.00620},
year={2021}
}
Abstract
This project investigates the effects of five classical 2D attention modules (Non-local, Criss-cross, Squeeze-Excitation, CBAM, Dual-attention) and five novel 3D attention modules (Attentional-ShapeContextNet, Point-Attention, Channle Affinity Attention, Offset-Attention, Point-Transformer) in 3D point cloud object detection, based on VoteNet pipeline.
Our Attentional Backbone
Results
Visualization
Settings
Set up the VoteNet project, and replace the models/backbone_module.py file with ours.
Trained-models
The trained models can be downloaded from google drive.
The detailed evaluation logs reported in the paper can be found at google drive.
Acknowledgment
The code is built on VoteNet. We thank the authors for sharing the codes.
