Depth estimation from monocular color images using natural scene statistics models

Abstract

We consider the problem of estimating a dense depth map from a single monocular image. Inspired by psychophysical evidence of visual processing in human vision systems (HVS) and natural scene statistics (NSS) models of image and range, we propose a Bayesian framework to recover detailed 3D scene structure by exploiting the statistical relationships between local image features and depth variations inherent in natural images. By observing that similar depth structures may exist in different types of luminance/chrominance textured regions in natural scenes, we build a dictionary of canonical range patterns as the prior, and fit a multivariate Gaussian mixture (MGM) model to associate local image features to different range patterns as the likelihood. Compared with the state-of-the-art depth estimation method, we achieve similar performance in terms of pixel-wise estimated range error, but superior capability of recovering relative distant relationships between different parts of the image.