SF300 large-scale image dataset
Samples of the SF300 dataset. Contains data from Styrelsen for dataforsyning og effektivisering, "Skraafoto", Feb. 2019.
The SF300 dataset is a large-scale dataset dedicated to the training of deep descriptors for remote sensing tasks. With orientation variations and a class-definition based on geolocation rather than task-specific annotations, it is particularly suited for fine-tuning CNN feature extractors for remote sensing classification or retrieval. The dataset was made using images taken from a plane and covers a wide range of landscapes on the Danish territory.
It consists of 308k images in 27.5k unique geolocations (classes) for the train set, and 21.8k images in 2.4k classes for the test set that can be used as the val split when fine-tuning. Images are all 512x512.
Details and use rights
These data are made available for research purposes only. For more details on them and their conditions of use, please consult the following documents:
- Statistics on the dataset: SF300_stats.pdf
- Usage and rights: README.md
Download
Id: benchmarkAlegoria
Passwd: EePeeghehoZ7aeDa
We provide utilities for training and testing feature extractors on SF300 (PyTorch) here:
https://github.com/dgominski/RSFineTuning
Reference article
D. Gominski, V. Gouet-Brunet, and L. Chen, “Unifying Remote Sensing Image Retrieval and Classification with Robust Fine-tuning”, ArXiv:2102.13392 [Cs], 26 Feb. 2021. Available at http://arxiv.org/abs/2102.13392
Date of publication: 12/03/2021