Benchmarking-the-Robustness-of-Panoptic-Segmentation-for-Automotive-Cameras

This is the PyTorch re-implementation of our aiXiv paper: Benchmarking the Robustness of Panoptic Segmentation for Automated Driving, link.


• New robustness dataset: Degraded-Cityscapes+ (D-Cityscapes+) (New model for snow and unfavourable light)
• Unifying degradation pipeline: unifying pipeline to assess the robustness of panoptic segmentation models for assisted and automated driving (AAD) systems, correlating it with image quality.
• Benchmarking experiments: 3 state-of-the-art CNN- and transformer-based panoptic segmentation networks are used to compare their robustness.

Requirements

Here, we show examples of using the EfficeintPS, DeepLab, and Oneformer for the Cityscape dataset.

• Requirement for EfficientPS is from here.
• Requirement for DeepLab is from here.
• Requirement for Oneformer is from here.
• Download the Cityscape validation here.

More coming soon upon!

To-do List

• Share all the D-Cityscapes+ dataset with 19 noise factors
• Share the noise generation table and tools
• Share the re-trained model on the new noisy data under adverse weather conditions

Citation

If you find this code helpful in your research or wish to refer to the baseline results, please use the following BibTeX entry.

@article{wang2024benchmarking,
  title={Benchmarking the Robustness of Panoptic Segmentation for Automated Driving},
  author={Wang, Yiting and Zhao, Haonan and Gummadi, Daniel and Dianati, Mehrdad and Debattista, Kurt and Donzella, Valentina},
  journal={arXiv preprint arXiv:2402.15469},
  year={2024}
}