fMoW-RGB-baseline

This dataset is a preprocessed version of the Functional Map of the Word (fMoW) dataset.

The fMoW-RGB dataset version was preprocessed strictly following the official baseline code. In short, original images are cropped according to their AOIs, and resized to 224×224 pixels JPEGs. Associated to the images, we provide metadata JSON files with some fields processed such that they match the cropped-resized images (such as ground sampling distance and the geographic boundaries).

Data format

Each data sample follows following file naming scheme. <split>/<class>/<class>_<sequence-id>/<class>_<aoi-id>/<class>_<sequence-id>_<index>_rgb.<jpg-or-json>. Each sample consists in one image (JPEG) and one metadata file (JSON). The path without extension constitutes the sample ID.

Images are 8-bit 224×224 pixels compressed JPEGs.

The JSON metadata file format follow the original data. For metadata fields whose value may be different from the original, the original key-values are kept under the field named "orig". Metadata keys are best documented in the Christie et al. (2018) - Appendix I.

We provide the data archived across multiple tarballs named <split>-<images-or-metadata>.tar.gz. For instance, the train images are in train-images.tar.gz and metadata in train-metadata.tar.gz.

Data splits

The dataset splits are identical to the original fMoW splits. The splits provided are "train" and "val", identical to the original fMoW splits. The other splits of the original dataset ("test" and "seq") are not provided yet. The "train" split contains 363,571 samples and the "val" split contains 53,042 samples. The standard procedure is to use the "train" split for training and "val" split for validation and testing (as done in Ayush et al. (2021) and following works).

Text files with samples for the different splits are available in the splits directory. Each row contains a sample ID.

Downloading

To download and extract all the data, run download.sh.

Attribution

The fMoW dataset was proposed in Christie, Gordon, et al. "Functional Map of the World." Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2018. https://doi.org/10.1109/CVPR.2018.00646.

Licence

This dataset is adapted material from the fMoW dataset, and is licensed under the same terms, namely the FUNCTIONAL MAP OF THE WORLD CHALLENGE PUBLIC LICENSE.