SGAI25 Workshop - Applied XAI

This Repository contains code that has been presented at the SGAI 25. More information on https://www.bcs-sgai.org/ai2025/?section=workshops

To cite this, please use the publication

Manss, Christoph, and Tarek A. El-Mihoub. 
"Evaluation of Explanations for Object Detection Using Transformers with Sonar Data." 
International Conference on Innovative Techniques and Applications of Artificial Intelligence. 
Cham: Springer Nature Switzerland, 2025.

Quick setup

1. Create and activate a Python environment (recommended)

• Python 3.10+
• Then install dependencies:

pip install -r requirements.txt

2. Get the demo data (MarineDebris) — choose one:

• Manually from the release page: https://github.com/mvaldenegro/marine-debris-fls-datasets/releases/tag/watertank-v1.0
• Or run the helper script (downloads and unpacks into data/):
• Prior to executing the download_marineDebris.sh script, please install python3.12-venv for virtual environments

bash download_marineDebris.sh

3. Get pretrained model checkpoints into models/:

bash download_models.sh

This script downloads a tar archive from the provided OwnCloud link and unpacks it directly into models/.

Notes

• Requirements: bash, curl, and tar must be available (standard on Linux/macOS; on Windows use Git Bash or WSL).
• Both scripts are idempotent to the extent supported by tar extraction; re-running is safe and will overwrite existing files if present in the archive.

Running the workshop notebooks

Launch Jupyter and open the notebooks in the notebooks/ folder. Each notebook installs the project requirements in its first cell, so running top-to-bottom should work if the data and models folders are prepared as above.

Start Jupyter (example):

jupyter lab  # or: jupyter notebook

• notebooks/workshop_attentions_from_transformer.ipynb

  • Explores self- and cross-attention from transformer-based object detectors as explanations. Loads a batch, runs inference, and visualizes attention maps interactively.

• notebooks/workshop_D-Rise_on_Transformer.ipynb

  • Demonstrates D-RISE for object detection with transformer backbones. Generates saliency maps and visualizes detections plus explanations. Includes simple guidance on thresholds and parameters.

• notebooks/workshop_ODAM_on_Transformer.ipynb

  • Shows ODAM (Object Detector Activation Maps) on transformer detectors. Extracts intermediate features with hooks, computes activation-based saliency, and evaluates with localization/faithfulness metrics.

Tips

• If you have multiple checkpoints under models/**/checkpoints/, adjust the checkpoint path at the top of each notebook to the one you want to use.
• The notebooks use Hugging Face AutoImageProcessor for preprocessing and postprocessing; make sure requirements.txt has been installed.

Acknowledgements

The work presented in this repository is funded by the Federal Ministry of Education and Research, Germany, grant number 01IW23003.