Merrielle Spain

Measuring Saliency and Frequency of Objects in Our Visual World

How many objects do we recognize in images? How visually important are the objects relative to each other? How frequently do they appear? In what context? Answering these questions will help us design better visual recognition algorithms by providing design constraints, performance specifications as well as ground truth for experiments. We designed a method for harvesting the names of visual categories seen by people in everyday pictures. Based on these data, as well as data collected by LabelMe, we develop and validated simple model of human object naming which enables us to propose a principled definition of saliency. Furthermore, we investigate the number of visual categories that are present in well-identified human environments, the frequency with which these categories occur, as well as their correlation.