Evaluating feature attribution methods in the image domain

Feature attribution maps are a popular approach to highlight the most important pixels in an image for a given prediction of a model. Despite a recent growth in popularity and available methods, the objective evaluation of such attribution maps remains an open problem. Building on previous work in this domain, we investigate existing quality metrics and propose new variants of metrics for the evaluation of attribution maps. We confirm a recent finding that different quality metrics seem to measure different underlying properties of attribution maps, and extend this finding to a larger selection of attribution methods, quality metrics, and datasets. We also find that metric results on one dataset do not necessarily generalize to other datasets, and methods with desirable theoretical properties do not necessarily outperform computationally cheaper alternatives in practice. Based on these findings, we propose a general benchmarking approach to help guide the selection of attribution methods for a given use case. Implementations of attribution metrics and our experiments are available online (https://github.com/arnegevaert/benchmark-general-imaging).