How do different sources of information arising from different modalities interact to control where we look? To address this question with respect to real-world operational conditions we presented natural scenes to human subjects in a variety of conditions and measure subjects' eye-movements. Using quantitative measures we find that (1) the influence of top-down signals is systematically stronger than that of spatial biases, which in turn is stronger than bottom-up signals; (2) the information is optimally integrated along a scanning trajectory; (3) the empirical saliency is combined linearly; (4) the action of top-down information can not be explained by a modulation of bottom-up signals. (5) In spite of these constraints, a large scale combination of features in a bottom-up driven model can explain a significant part of the variability of eye movements as measures by an ROC analysis (AUC=0.75).