BEV-MAE: Bird's Eye View Masked Autoencoders for Point Cloud Pre-training in Autonomous Driving Scenarios
This is the official implementation of BEV-MAE.
We release the pre-training weights of VoxelNet on Waymo dataset.
| pre-trained 3D backbone | Dataset | Weights |
|---|---|---|
| VoxelNet | Waymo (20% data) | Google_drive |
| VoxelNet | Waymo (full data) | Google_drive |
Our code is base on OpenPCDet (0.5 version). To use our pre-trained weights, please refer to INSTALL.md for installation and follow the instructions in GETTING_STARTED.md to train the model.
See the scripts in tools/run.sh
BEV-MAE is based on OpenPCDet. It is also greatly inspired by the open-source code Occupancy-MAE.
If BEV-MAE is useful or relevant to your research, please kindly recognize our contributions by citing our paper:
@inproceedings{lin2024bevmae,
title={BEV-MAE: Bird's Eye View Masked Autoencoders for Point Cloud Pre-training in Autonomous Driving Scenarios},
author={Lin, Zhiwei and Wang, Yongtao and Qi, Shengxiang and Dong, Nan and Yang, Ming-Hsuan},
booktitle={Proceedings of the AAAI conference on artificial intelligence},
year={2024}
}If you have any problem about this work, please feel free to reach us out at zwlin@pku.edu.cn.
The project is only free for academic research purposes, but needs authorization for commerce. For commerce permission, please contact wyt@pku.edu.cn.