Single image depth estimation using local gradient-based features

The paper considers single image depth estimation for urban outdoor content images. The proposed approach uses supervised machine learning to learn the relationships between low level image features and depth gradient. Brightness, color and texture gradient cues are used as features. Markov Random Field (MRF) model is employed to estimate depth gradient and the model parameters are learned trough linear regression. The depth gradient in horizontal and vertical direction is modeled independently. The estimation of the depth gradient is preformed using Maximum A Posteriori Probability (MAP) estimation. The final depth map for an image is calculated by integrating the estimated depth gradients. The experimental results show that the approach based on relatively simple model achieves very good results for urban outdoor images.