XAI-project

Quantitatively evaluating explainability: Benchmarking XAI Techniques for image classification tasks

This project benchmarks explainability methods for binary image classification using semantically grounded regions extracted with Grounded-SAM.

We evaluate both gradient-based (e.g., Integrated Gradients, Grad-CAM) and perturbation-based (e.g., LIME, Kernel SHAP) methods over the same interpretable input space.

Main Contributions

• Use of semantic masks (e.g., head, ears, tail) instead of superpixels or raw pixels
• Fair evaluation across multiple XAI methods and CNNs (ResNet18, VGG16, MobileNetV2)
• Quantitative comparison using:
  • Effective Compactness
  • Rank Quality Index (RQI)
  • Stability
  • Sufficiency Score
  • Valley Score
  • Execution Time

Image Example of patches working for Sufficiency Score (Occlusion, ResNet-18 , Cats and Dogs dataset).

Datasets

• Cars vs. Bikes
• Cats vs. Dogs

Notebooks

• single_sample_Resnet18_CatsDogs.ipynb: full pipeline example on one image
• A base implementation using standard SAM, which does not require human input is available at /others/SAM_starting_pipeline.ipynb

This project is part of a research paper on benchmarking XAI for visual binary classification.

Developed by Mattia Viglino, Vincenzo Montana and Anna Lisa Maddaloni as part of the Explainable AI course at Politecnico di Torino 2025.