Predicting human gaze using low-level saliency combined with face detection (W49)
Moran Cerf, Jonathan Harel, Wolfgang Einhäuser, Christof Koch
low-level saliency

+ =
fixations
face detection fixation prediction

Humans look at faces (early and irrespective of task)

Combined Model (face + low-level saliency)

- Combined model outperforms saliency in predicting fixations. - Holds even if there is no face in the image (hallucinated faces). => Specialized bottom-up "face channel" for human attention.