Personalized saliency prediction using color spaces

Saliency is the ability of being important, noticeable or attention worthy. Finding salient regions in images has important applications in automatic image cropping, image compression and advertisements. The salient regions for an individual in an image changes according to their gender, race, culture, likes, dislikes and experiences. Universal Saliency Maps point out the overall general salient regions without any considerations of personal traits of the subject. Therefore, personalized saliency maps are required for better and more personalized predictions of salient regions. In this study, using the RGB (Red, Green, Blue), CYMK (Cyan, Yellow, Magenta, Key), HSV (Hue, Saturation, Value) and HSL (Hue, Saturation, Lightness) fixation patterns of individuals, we propose a Gradient Boosted Tree Regression model to extract personalized saliency map from the universal saliency map with an average accuracy of 80% (Area Under Curve Judd Metrics). We also put forth our discussion for why some images and subjects have better saliency map predictions than others.