DAOccV2

DAOccV2: Efficient Fusion and Occlusion-Aware Augmentation for Multi-Modal 3D Occupancy Prediction

Zhen Yang, Pengyi Guo, Jiayu Wang, Yanpeng Dong, Jiwen Lu

DAOccV2 is an efficient camera-LiDAR 3D occupancy prediction framework built upon DAOcc. It improves inference efficiency by skipping dense image feature lifting for empty voxels and enhances foreground occupancy prediction through occlusion-aware ground truth augmentation and instance-level auxiliary supervision.

Highlights

• Efficient and accurate. On Occ3D-nuScenes with camera masks, DAOccV2 reaches 56.47 mIoU at 8.3 FPS.

News

• 2026-07-10: Release the model weights and the first version of the code.

Experimental Results

Occ3D-nuScenes

Model	Training setting	mIoU	RayIoU	mIoUF	FPS	Config	Weights
DAOccV2	w/ camera mask	56.47	-	55.53	8.3	config	weights
DAOccV2	w/o camera mask	50.6	50.4	52.2	8.3	config	weights

SurroundOcc

Model	IoU	mIoU	mIoUF	Config	Weights
DAOccV2	45.4	32.7	31.2	config	weights

Occ3D-Waymo

Model	Training setting	mIoU	mIoUF	Config	Weights
DAOccV2	w/ camera mask	47.38	39.30	config	-

• Weights trained on Waymo cannot be shared per Waymo’s regulations.
• Following prior works, the Occ3D-Waymo experiment uses 20% of the training data.
• $mIoU_F$ denotes the mean IoU over foreground classes.

Getting Started

• Installation
• Prepare datasets
• Training and evaluation
• Visualization

Citation

Citation information for DAOccV2 will be added when the paper is publicly available. If this codebase or its DAOcc foundation is useful in your research, please also cite:

@article{yang2025daocc,
  title={Daocc: 3d object detection assisted multi-sensor fusion for 3d occupancy prediction},
  author={Yang, Zhen and Dong, Yanpeng and Wang, Jiayu and Wang, Heng and Ma, Lichao and Cui, Zijian and Liu, Qi and Pei, Haoran and Zhang, Kexin and Zhang, Chao},
  journal={IEEE Transactions on Circuits and Systems for Video Technology},
  year={2025},
  publisher={IEEE}
}

Acknowledgements

This project builds upon these excellent open-source codebases:

• MMDetection3D
• DAOcc