Extracting image features in static images for depth estimation

Human feels three-dimensional effect for static image with the cues of various kinds of image features such as relative sizes of objects, up and down, rules of perspective, texture gradient, and shadow. The features are called pictorial depth cues. Human is thought to learn to extract these features as important cues for depth estimation in the developmental process. In this paper, we make a hypothesis that pictorial depth cues are acquired so that disparities can be predicted well and make a model that extracts features appropriate for depth estimation from static images. Random forest network is trained to extract important ones among a large amount image features so as to estimate motion and stereo disparities. The experiments with simulation and real environments show high correlation between estimated and real disparities.