Comparison of different methods

>Failed samples:

>Explanation:

-In the first to the third row, our algorithm produces unsatisfactory saliency maps due to the false positives or false negatives of the face detector.
-In the images in the forth to the sixth row, our algorithm detects the important faces, but misses other salient regions such as the hat in the forth row or the canvas in the sixth row. However, like us, the other algorithms either miss the important faces or fail to detect the other salient objects.
-In the images of the seventh and eighth row, the most salient objects lie near the boundary. Due to the center prior parameter in our algorithm, we miss these salient objects. The other algorithms that incorporate center priors also fail.
-For the images in the ninth row, observers attribute high saliency to the blurry building in the background. However, our algorithm attributes saliency to the foreground crowd.
-For the images in the tenth row, people do not indicate any specific salient object, as shown by the ground truth. All the algorithms produce fuzzy saliency maps over the whole image anyway, as there is plenty of local contrast in the image.