Scanpaths analysis with fixation maps to provide factors for natural gaze control

Abstract

Human like gaze control is needed for robots to be companions of humans. For human-like gaze control, much research has been progressing to identify factors that affect human gaze. The conventional approaches to discover factors that affect human gaze is based on their hypotheses. They presented gaze control algorithm based on hypotheses and verified through experiments. However, since the algorithms were originated from the hypotheses, they were prone to be biased to the hypotheses. This paper derives the factors that affect human gaze based on observation of real humans scanpaths, not hypothesis. From the recorded scanpaths, fixation maps are produced using the Gaussian distribution. The earth movers distance (EMD) is used to measure similarity among fixation maps, and the fixation map with minimal difference is selected in each test image. The selected fixation maps are used to derive the factors that affect human gaze. The derived factors are center, salient regions, human, depth, objects, scene schema information, and they are shown with examples.