Welcome to the official website of the Dronescapes dataset. This dataset aims to provide a comprehensive starting point for teaching machine learning models various tasks related to Aerial Image Understanding using cameras from UAVs. To do so, it defines three main tasks: semantic segmentation, metric depth estimation and camera normals estimation.
The original dataset was introduced in our paper accepted at ICCV 2023: link. For more details about this particular dataset, please see its dedicated page: link.
Later, in 2025 the dataset was further extended with new videos and modalities, resulting Dronescapes 2 & 3. For technical details about this dataset, which builds on top of the first one, see its dedicated page: link.
The Dronescapes dataset provides an official benchmark which should be used for a fair comparison between methods on a fixed subset of frames. It is called Dronescapes-Test and was introduced in the first iteration of the Dataset. For a comprehensive method of computing the metrics, up to date state-of-the-art results, as well as details regarding the test set, please see this page: link.
The actual files of this dataset are hosted on Huggingface: Dronescapes (2023) and Dronescapes 2 & 3 (2025). Follow the instructions from there on how to download the data, load the files and start training your own models!
To cite our work, please use the following: