Quantitative evaluation on background-only images (no salient objects)

Develop quantitative evaluation metrics for salient object detection on images that contain no salient objects (i.e., background-only images), where standard measures such as precision–recall, ROC/AUC, and F-Measure are inapplicable due to empty positive ground-truth labels and mean absolute error is uninformative under normalized saliency maps. Design and validate measures that properly assess false positive suppression and map quality in the absence of positive annotations.

Background

The paper highlights that nearly all salient object detection models assume the presence of at least one salient object, which is impractical for background-only images comprising textures or clutter. On such images, good models should produce dark, non-activated saliency maps, but leading methods often respond to background clutter due to reliance on contrast.

The authors explain that conventional quantitative metrics fail in this setting: PR and ROC (and thus F-Measure and AUC) cannot be computed because the positive ground-truth set is empty, and MAE is not informative since many methods normalize saliency maps. This gap motivates the need for new, specific evaluation criteria for background-only cases.