VOGS-CP

[AAAI 2026 Oral] Vision-Only Gaussian Splatting for Collaborative Semantic Occupancy Prediction

Paper | Project page

Prepare Datasets

• OPV2V: Download from HERE.
• Semantic-OPV2V: Please refer to CoHFF 4LidarSurround version.

Organize as follows:

VOGS-CP/dataset

. 
├── OPV2V
│   ├── surround
│   ├── test
│   ├── train
│   └── validate

or ln -s /path/to/opv2v dataset/OPV2V

Prepare Environments

• OpenCOOD: Please refer to HEAL.

Additionally,

pip install openmim
mim install mmcv==2.1.0
mim install mmdet==3.3.0
mim install mmsegmentation==1.2.2

# deformable attention & gaussian-to-voxel splatting
(cd opencood/models/gaussian_modules/ops && pip install -e .)
(cd opencood/models/gaussian_modules/localagg && pip install -e .)

Quick Start

Please refer to HEAL to get familiar with Basic Train / Test Command.

Example Commands

Download the pretrained weights for the image backbone HERE and put it inside ckpts.

# Single-agent training
mkdir opencood/logs/m4_single_25600
cp opencood/hypes_yaml/opv2v/gaussian/m4_single_25600/config.yaml opencood/logs/m4_single_25600
python opencood/tools/train.py -y None --model_dir opencood/logs/m4_single_25600

# Collaborative training
mkdir opencood/logs/m4_collab_25600_0.4
cp opencood/hypes_yaml/opv2v/gaussian/m4_collab_25600_0.4/config.yaml opencood/logs/m4_collab_25600_0.4
cp opencood/logs/m4_single_25600/your/bestval/checkpoint opencood/logs/m4_collab_25600_0.4/net_epoch1.pth
python opencood/tools/train.py -y None --model_dir opencood/logs/m4_collab_25600_0.4

# Collaborative inference
python opencood/tools/inference_heter_task.py --model_dir opencood/logs/m4_collab_25600_0.4 --task occupancy --range 20,20

Acknowledgements

Our implementation benefits from a lot of awesome previous works, such as: STAMP, HEAL, GaussianFormer, CoHFF.

Citation

If you find this project helpful, please consider citing the following paper:

@article{chen2025vision,
  title={Vision-Only Gaussian Splatting for Collaborative Semantic Occupancy Prediction},
  author={Chen, Cheng and Huang, Hao and Bagchi, Saurabh},
  journal={arXiv preprint arXiv:2508.10936},
  year={2025}
}