SemanticAdaBins - Using Semantics to Improve Depth Estimation based on Adaptive Bins in Aerial Scenarios

Supervised depth estimation from a single image (Monocular Depth Estimation - MDE) is particularly difficult, especially in aerial scenarios. A major challenge in these unstructured environments is determining the optimal depth discretization to accurately map depth points. Finding the proper bin for each depth is thus crucial for significantly improving both the accuracy and effectiveness of depth estimation. In this work, we introduce a novel approach to integrate semantic information into the binning procedure, inherent to the top methods in supervised monocular depth estimation. Our new method extends adaptive bins by separating the scene according to semantic classes and customizing bins for each semantic class. This is done by employing a novel visual transformer that uses semantic-aware embeddings to generate bin widths for each semantic class. Results on Uavid aerial images show the effectiveness of the approach.