Monocular Depth Prediction with a Fully Convolutional Neural Network and Skip Connections.

This work tackles the problem of estimating depth from a single RGB image of a scene. To model the complex relationship between depth and monocular images, we propose a fully convolutional neural network that incorporates skip connections along with encoding and decoding stages to output consistent and detailed depth maps. Additionally, our model is a single convolutional architecture that does not use post-processing strategies; thus, it equires relatively less computational power and time to train when compared to more complex works. Experimental analysis and evaluation with a variety of losses and accuracies show that the end-to-end training process results in a model that performs better than or similar to many past architectures trained on the same dataset. [ABSTRACT FROM AUTHOR]