Investigating Attention Mechanism in 3D Point Cloud Object Detection (3DV 2021)

This repository is for the following paper: "Investigating Attention Mechanism in 3D Point Cloud Object Detection"
International Conference on 3D Vision (3DV 2021)
Shi Qiu*, Yunfan Wu*, Saeed Anwar, Chongyi Li

Paper and citation

The paper can be downloaded from here (arXiv) and here (ieee).
If you find our paper/codes/results are useful, please cite:

@inproceedings{qiu2021investigating,
  title={Investigating Attention Mechanism in 3D Point Cloud Object Detection},
  author={Qiu, Shi and Wu, Yunfan and Anwar, Saeed and Li, Chongyi},
  booktitle={2021 International Conference on 3D Vision (3DV)},
  pages={403-412},
  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.