Paper: IEEE Xplore, arXiv

We list our environment setup below:

• Python 3.8
• PyTorch 1.12.0+cu113
• MMCV 1.6.0
• MMDetection 2.25.0
• MMSegmentation 0.26.0
• MMDetection3D 1.0.0rc3
• vod-tudelft 1.0.3 (This is the toolkit of vod dataset, and can be installed by pip install vod-tudelft==1.0.3 command.)

After setting up the environment, please move the files in this repo to your mmdetection3d folder.

Please use the file provided in tools/create_data_vod.py to generate the corresponding data.

python tools/create_data_vod.py --root-path ${YOUR_DATA_PATH}$

Please also make sure you edit the data_root in plugin/lxl/configs/_base_/datasets/vod_r_c_3classes.py to point to the correct data directory.

Please download the pretrained model in MMDetection Model Zoo (the YOLOX-s model in https://github.com/open-mmlab/mmdetection/tree/main/configs/yolox) and make sure you edit the pretrained_img term in plugin/lxl/configs/lxl/LXL_vod.py to point to the correct directory.

To train LXL with a single GPU, you can use the following command:

python tools/train_v2.py plugin/lxl/configs/lxl/LXL_vod.py

To evaluate the trained model, you can use the following command:

python tools/test_v2.py plugin/lxl/configs/lxl/LXL_vod.py ${YOUR_CHECKPOINT_PATH}$ --eval bbox

You can download our trained model here. Note that we have refactored our code and trained a new model, so that the performance is slightly different from that reported in our paper.

@ARTICLE{xiong2024lxl,
  author={Xiong, Weiyi and Liu, Jianan and Huang, Tao and Han, Qing-Long and Xia, Yuxuan and Zhu, Bing},
  title={LXL: LiDAR Excluded Lean 3D Object Detection With 4D Imaging Radar and Camera Fusion}, 
  journal={IEEE Transactions on Intelligent Vehicles},
  volume={9},
  number={1},
  pages={79-92},
  year={2024},
  doi={10.1109/TIV.2023.3321240}
}