WTA/TLA: A UAV-captured Dataset for Semantic Segmentation of Energy Infrastructure

Automated inspection of energy infrastructure with Unmanned Aerial Vehicles (UAVs) is becoming increasingly important, exhibiting significant advantages over manual inspection, including improved scalability, cost/time effectiveness, and risks reduction. Although recent technological advancements enabled the collection of an abundance of vision data from UAVs’ sensors, significant efforts are still required from experts to interpret manually the collected data and assess the condition of energy infrastructure. Thus, semantic understanding of vision data collected from UAVs during inspection is a critical prerequisite for performing autonomous robotic tasks. However, the lack of labeled data introduces challenges and limitations in evaluating the performance of semantic prediction algorithms. To this end, we release two novel semantic datasets (WTA and TLA) of aerial images captured from power transmission networks and wind turbine farms, collected during real inspection scenarios with UAVs. We also propose modifications to existing state-of-the-art semantic segmentation CNNs to achieve improved trade-off between accuracy and computational complexity. Qualitative and quantitative experiments demonstrate both the challenging properties of the provided dataset and the effectiveness of the proposed networks in this domain.The dataset is available at: https://github.com/gzamps/wta_tla_dataset.