Vision-Based Panoptic Occupancy Prediction in Urban Environments

This repository contains the implementation of our paper.

Getting Started

• Installation

• Prepare data

• Train

  # Single GPU
  python3 tools/train.py ./configs/sfmocc/sfmocc.py --panoptic
  
  # 8 GPUs
  ./tools/dist_train.sh ./configs/sfmocc/sfmocc.py --panoptic 8

• Evaluation

  # Single GPU
  python3 tools/test.py ./configs/sfmocc/sfmocc.py ./path/to/ckpts.pth --panoptic
  
  # 8 GPUs
  ./tools/dist_test.sh ./configs/sfmocc/sfmocc.py ./path/to/ckpts.pth --panoptic 8

• Visualization

  # Save predictions and images (select scene-id)
  python3 tools/test.py configs/sfmocc/sfmocc.py ./path/to/ckpt.pth --dump_dir=pred_dir --scene xxxx
  
  # Generate video (select scene-id)
  python3 tools/visualization/visual.py pred_dir/scene-xxxx

Acknowledgement

Many thanks to the authors of RenderOcc for the codebase.

Citation

@article{marcuzzi2026icra,
  title={},
  author={},
  journal={},
  year={2026}
}

Licence

Copyright 2026, Rodrigo Marcuzzi, Cyrill Stachniss, Photogrammetry and Robotics Lab, University of Bonn.

This project is free software made available under the MIT License. For details see the LICENSE file