A methodology to evaluate the effectiveness of saliency map generation methods

In the Explainable Artificial Intelligence (XAI) field, many post-hoc visual explanation techniques for generating saliency maps have been proposed to explain computer vision downstream task output. However, each new method is often presented as the best approach to provide these explanations despite having been tested on a limited number of scenarios, generating confusion or presenting contradictory results with other similar work. Therefore, we present a comprehensive and unbiased comparison of various XAI techniques for generating saliency maps based on the image classification downstream task. The aim is to determine which method offers the best explanations and evaluate them from different perspectives. We consider factors such as training strategy, domain adaptation, performance in fine-grained or unbalanced problems, and scalability with diverse classes and samples.