Geometry-Aware Spatio-Temporal Context Modeling for 4D Occupancy Forecasting (ECCV 2026)

June 24, 2026 ยท View on GitHub

Environment

  1. Create the conda environment
conda create -n occ python=3.7
conda activate occ
  1. Install the following dependencies
pip install torch==1.12.1+cu116 torchvision==0.13.1+cu116 torchaudio==0.12.1 --extra-index-url https://download.pytorch.org/whl/cu116
pip install torch-scatter -f https://data.pyg.org/whl/torch-1.12.0+cu116.html
pip install mmcv-full -f https://download.openmmlab.com/mmcv/dist/cu116/torch1.12.1/index.html
pip install spconv-cu116 tensorboard numba einops pyquaternion nuscenes-devkit

Prepare Dataset

  1. Download Occ3D-nuScenes from Occ3D-nuScenes.

  2. Download the generated info file from Google Drive.

Model Zoo

MethodInputEgo traj.mIoU (%)IoU (%)Weights
GAST3D-OccPred.27.3835.90Google-drive
GAST3D-OccGT47.4056.24Google-drive

Training and Evaluation

  1. Modify the configuration file config/gast_nuscenes.yaml.

  2. Enter the tasks/occ3d_nuscenes folder, modify run.sh and run the command ./run.sh.

Acknowledgements

Thanks to the excellent projects: OccWorld, II-World