Joint Estimation of Depth and Geometric Scene Properties From Single Images

Monocular depth estimation is the task of obtaining a measure of distance for each pixel using a single image. It is an important problem of computer vision that may be solved with neural networks. Though recent works in this area have shown significant improvement in accuracy, state-of-the-art methods require large memory and time resources. The primary purpose of this work is to improve the performance of the latest solutions with no decrease in accuracy. To achieve this, we propose a Double Refinement Network architecture along with linear depth regularization. Training process uses multiple datasets to supervise depth estimation and line segment detection tasks. Using the proposed method, the results on standard benchmark RGB-D dataset NYU Depth v2 achieve state-of-the-art quality while frames per second rate of our approach is significantly higher, and RAM per image is lower.