Computational Models of Salience and Task on Visual Search

Human eye-movements during visual search a guided both with bottom-up and top-down cues. In the present thesis I improve bottom-up salience model (Itti & Koch, 2001) by adding a semantical information to it. This top-down component uses Latent Dirichlet Allocation (LDA) algorithm which has been suggested to be a model of semantic memory.  LDA was trained on a dataset of annotated images (Russell et. Al, 2008), and then used to build a priority map showing the semantics relatedness of the image objects to the target object. The priority map is merged with the saliency map from  (Itti & Koch, 2001) resulting in an importance map for an integrated model. I conducted an eye-tracking study where participants searched for “cup” and “painting” on a set of  natural indoor images. The results of the experiment show that integrated importance model is significantly (p<0.001) different from the saliency model and is 5% better in predicting human fixations for “cup” target. It did not improved saliency model for the painting target because painting is more general object in terms of visual semantics.