Boosting bottom-up and top-down visual features for saliency estimation