Depth estimation using monocular cues from single image

Abstract

This paper investigates depth estimation using monocular cues. Human visual system uses monocular cues such as texture, focus and shading for depth perception. Our proposed algorithm is based on segmenting the image into homogenous segments (superpixels), and then out of these segments we extract the ground segment and the sky segment. These two segments guide the depth estimation by providing region with maximum depth (sky) and region with minimum depth (ground). The reset of the segments will have a depth value between the sky and ground. This algorithm address image that contains sky and ground as a part of the image. The ground acts as a support for segments (eg. Trees, buildings) in the image, thus a vertical image segments tends to have similar depth as its ground support. On the other hand, some images are not supported by the ground but they are connected to it, therefore these segments will have depth value larger than its nearest ground pixels.