Dataset Card for LAION-EO

Dataset Summary

This dataset contains a subset of LAION-5B containing images that are likely to be satellite images. The procedure of acquiring and filtering the dataset has been described in https://arxiv.org/abs/2309.15535.

Version	Number of Samples
0	24,933
1	112,985

Dataset Structure

Each version of the dataset contains a .csv file with metadata with urls to images, which can be easily filtered. Note that the linked images could be copyrighted.

Data Fields

Field	Description
source	Index of the anchor sample
url	Link to the image
filename	Locally saved unique filename
id	Original ID
fast_similarity	Fast similarity to the anchor image computed with https://github.com/rom1504/clip-retrieval
caption	Text caption
image_similarity	CLIP similarity to the original anchor image
text_similarity	CLIP similarity to the text "a satellite image"
height	height of the image at url
width	Width of the image at url
lang	Language predicted using https://huggingface.co/papluca/xlm-roberta-base-language-detection
lang_score	A measure of confidence in the predicted language

Example Samples

Data Splits

No official splitting of the dataset is used.

Dataset Creation

The creation of the prototype version is described in https://arxiv.org/abs/2309.15535.

Curation Rationale

Extraction of samples in LAION-5B relevant to Earth observation tasks.

Source Data

Samples from the existing LAION-5B dataset (https://laion.ai/blog/laion-5b/).

Discussion of Biases

Only contains satellite images openly uploaded online, which introduces a heavy bias towards satellite images used for communicating ideas on the internet.

Citation Information

The workshop paper presented at the DataComp workshop during ICCV 2023 is available at https://arxiv.org/abs/2309.15535.

@inproceedings{LAION_EO,
  title={From LAION-5B to LAION-EO: Filtering Billions of Images Using Anchor Datasets for Satellite Image Extraction}, 
  author={Mikolaj Czerkawski and Alistair Francis},
  year={2023},
  eprint={2309.15535},
  archivePrefix={arXiv},
  primaryClass={cs.CV}
  booktitle = {"Towards the Next Generation of Computer Vision Datasets: DataComp Track" Workshop at the IEEE/CVF International Conference on Computer Vision (ICCV)}
}

License

We distribute the metadata dataset (the parquet files) under the Creative Common CC-BY 4.0 license, which poses no particular restriction. The images are under their copyright.

Contributions

Design and Curation: Mikolaj Czerkawski