This is the implementation of the work: RoadBEV: Road Surface Reconstruction in Bird’s Eye View.

Road surface reconstruction, an essential task for autonomous driving comfort control, suffers from poor performance in perspective view. We propose RoadBEV-mono and RoadBeV-stereo, which reconstruct road surface with monocular and stereo images, respectively.

We use Road Surface Reconstruction Dataset (RSRD). The GT elevation maps should be generated off-line, as it costs much computation.

• First, download RSRD (the dense subset of 15.1GB) and the calibration files in dev kit. Set the corresponding path in dataset.py.

• Generate and save the GT maps to local. The GT maps depend on ROI and horizontal resolution. We do not recommend changing the related parameters in dataset.py, as they are determined based on statistic analysis on the dataset. Changes on them may lead to errors and inconsistent results. If you do change, re-generate corresponding GT.

python preprocess_gt.py --save_dir '/data/preprocessed/' --dataset 'train'
python preprocess_gt.py --save_dir '/data/preprocessed/' --dataset 'test'

The elevation maps along with masks are saved in .pkl format.

• Python 3.7
• Pytorch 1.10.0

conda create -n roadbev python=3.7
conda activate roadbev

conda install pytorch=1.10.0 torchvision torchaudio cudatoolkit=11.1 -c pytorch -c nvidia
conda install -c open3d-admin open3d
pip install opencv-python
pip install pillow

Use the following command to train RoadBEV-mono:

python train.py --lr 8e-4

Use the following command to train RoadBEV-stereo:

python train.py --stereo --lr 5e-4

python test.py --loadckpt 'xxx.ckpt'  # test RoadBEV-mono
python test.py --stereo --loadckpt 'xxx.ckpt'  # test RoadBEV-stereo

RoadBEV-mono, RoadBEV-stereo

If you find this project helpful in your research, welcome to cite the paper.

@misc{zhao2024roadbev,
      title={RoadBEV: Road Surface Reconstruction in Bird's Eye View}, 
      author={Tong Zhao and Lei Yang and Yichen Xie and Mingyu Ding and Masayoshi Tomizuka and Yintao Wei},
      year={2024},
      eprint={2404.06605},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}