Yuan Wu^1*, Zhiqiang Yan^2*, Yigong Zhang^3†, Xiang Li³, Jian Yang^1†

^*equal contribution    ^†corresponding author   
¹Nanjing University of Science and Technology    ²National University of Singapore    ³Nankai University   

Step1. Prepare environment as that in Install.

Step2. Prepare nuscenes and generate pkl file by runing：

python tools/create_data_bevdet.py

The final directory structure for 'data' folder is like

└── data
  └── nuscenes
      ├── v1.0-trainval
      ├── maps  
      ├── sweeps  
      ├── samples
      ├── gts
      ├── bevdetv2-nuscenes_infos_train.pkl 
      └── bevdetv2-nuscenes_infos_val.pkl

# train:
tools/dist_train.sh ${config} ${num_gpu}

# test:
tools/dist_test.sh ${config} ${ckpt} ${num_gpu} --eval mAP

The pretrained weights in 'ckpt' folder can be found here. All model weights can be found here.

This project builds upon several outstanding open-source projects. We sincerely thank the authors of:

• BEVDet, FlashOcc, FB-BEV, SCI, Occ3D, RoboBEV

If our method proves to be of any assistance, please consider citing:

@article{wu2025see,
  title={See through the Dark: Learning Illumination-affined Representations for Nighttime Occupancy Prediction},
  author={Wu, Yuan and Yan, Zhiqiang and Zhang, Yigong and Li, Xiang and Yang, Jian},
  journal={arXiv preprint arXiv:2505.20641},
  year={2025}
}