Human Visual Scanpath Prediction Based on RGB-D Saliency

Rui Han, Shuangjiu Xiao

Abstract: Human visual perception is considered as a dynamic process of information acquisition, while the visual scanpath can clearly reflect the shift of our eye fixations. In the previous study of visual attention, researchers generally do the saliency computation to predict where the regions of interest locate in the given scene, whereas less considering how our eyes saccade during the saliency generation. In this paper, we propose a novel model based on visual attention mechanism to predict the human visual scanpath of the given 3D scene. Our scanpath prediction model that can reasonably estimate the sequence of eye fixations when eyes saccade contains three important factors: RGB-D saliency computation, oculomotor biases and inhibition of return(IOR). In addition, we construct a small RGB-D eye tracking dataset with collecting eye tracking records from 91 people on 30 RGB-D images for our comparison experiments. The experiments demonstrate that our approach provides better prediction on human visual scanpath.