Monocular depth estimation with geometrical guidance using a multi-level convolutional neural network

Depth estimation using monocular images is highly challenging but is a considerable topic in understanding scene structure. This paper proposed a multi-level convolutional neural network (CNN), in which the low-level and high-level features were well-integrated in order to estimate the depth values from a single image. To estimate the depth values, a fully convolutional architecture which used a structure improvement strategy was applied to correct the depth values using low-level features of the shallow layers. A hierarchical context aggregation scheme was proposed according to the dilated convolutional operators that integrate the global and local contexts in a progressive way to recover the local details. In addition, a rectifying block was used to subtract the existed fitting residuals of integrated multi-level features. In the second level, a modifier network was provided to improve the estimated depth values, in particular in the object boundaries. It is possible to determine the objective function in the modifier networks by considering the geometrical features directing the network to achieve the best results. The proposed framework was evaluated using the computer vision (Make3D, NYU, and SUN datasets) and remote sensing (Vaihingen and Potsdam datasets) datasets that specified the significant performance of the proposed framework.