This chapter focuses on how to validate the attention prediction models based on ground truth coming from eye-tracking or mouse-tracking techniques. This ground truth is split into three kinds: (1) “object detection” models which will output the most salient object(s) in an image, (2) “eye-tracking ground truth” models which will provide static saliency maps as output, and the (3) “scan-path ground truth” models which provide a dynamic set of fixations through time. Those different outputs cover the existing attention models in the literature. For object detection validation, all the metrics are based on the notion of true or false positives or negatives. For eye-tracking ground truth, there are dozens of metrics (amplitude-based, location-based, distribution-based). For scan-path ground truth, there are also dozens of metrics (distances, density, vector/time series, recurrence). However, newer metrics try to unify scan-path ground truth and eye-tracking ground truth by creating time-evolving saliency maps instead of scan paths. Another new research track is about a fair fixation time duration validation as some new models can provide fixation locations but also fixations’ durations.