Edge Enhancement in Monocular Depth Prediction

Although many monocular depth prediction methods have achieved very high prediction accuracy, the preference of using high-level features of images makes these methods wrongly predict the depth of edge regions. This shortage does not decrease prediction accuracy seriously but will bring difficulties to subsequent works like three-dimension recognition and semantic segmentation. To enhance the performance of restoring the depth of edge regions, we apply modification on network structure and design a new loss function to strengthen the network's ability to extract, store, and utilize low-level features of images. We test our method on NYU Depth V2 Dataset, and the experiment results show that our method has a better performance on predicting the depth of edge regions than the state-of-the-art method and outperforms most of the current method on prediction accuracy.