A Critical Evaluation of Aerial Datasets for Semantic Segmentation

Drone perception systems use information from sensor fusion to perform tasks like object detection and tracking, visual localization and mapping, trajectory planning, and autonomous navigation. Applying these functions in real environments is a complex problem due to three-dimensional structures like trees, buildings, or bridges since the sensors (usually cameras) have limited viewpoints. We are interested in creating an application that is aimed towards inspection of forests with a focus on deforestation, with the main objectives being building a 3D semantic map of the environment and visual inspection of trees. In this paper, we evaluate three new datasets recorded at various flight altitudes, in terms of class balance, training performance on the semantic segmentation task, and the ability to transfer knowledge from one set to another. Our findings showcase the strengths of these datasets, while also pointing out their shortcomings, and offering future development ideas and raising research questions.