Evaluation of explainable AI localisation performance using relevance F-score

Explainable AI (XAI) has gained significant attention in the AI research community. However, the absence of a standard method for measuring the quality of explanations, and the diversity of existing XAI methodologies, make it hard to compare the performance of different XAI methods. An essential property of image-based XAI methods is their localisation performance, which describes how well the XAI method identifies the relevant object of interest. Existing methods for measuring XAI localisation performance do not consider the full XAI output, including image regions with negative saliency. To address the limitations of existing XAI localisation metrics, we introduce the Relevance F-score (RFS), which considers image regions with both positive and negative saliency to better capture an XAI method’s localisation performance. We evaluate our approach using a Visual Question Answering (VQA) network trained on CLEVR-XAI dataset.