saliency-benchmark

Repository for benchmarking different post-hoc XAI explanation methods on image datasets.

Install and Usage

To install and use the project, follow the step explained in the documentation.

For a rapid trial, you can find the extracted masks, the saliency maps computed for each method, the checkpoints of the trained models and the necessary occurrences for weight of evidence computation in the drive at the following link

Prediction and saliency map

ResNet model

Prediction	Image	GradCAM	LIME	RISE	SIDU
Golf ball
Glacier

VGG model

Prediction	Image	GradCAM	LIME	RISE	SIDU
Golf ball
Glacier

Alignment between human concepts and explainable deep learning models

Example of classification with VGG11 model, explained using GRADCAM as saliency method. The concepts extracted with Florence2 model of GroundedSAM2 are "Head" and "Paws".

Original Image	Saliency map	Concepts extracted	Saliency + Concepts