Extension with other RS captioned datasets

#1
by mespinosami - opened

Nice work! A direct extension to this is merging with other existing captioning EO datasets (e.g. https://github.com/201528014227051/RSICD_optimal).

Thanks! Definitely a good suggestion to train on multiple text-image datasets, and the one you linked looks like a good candidate.

Regarding merging, I would be somewhat reluctant to do that, since LAION-EO is focused on a specific type of data - images sourced from LAION-5B as it is exclusively derived from that. For that reason, I would rather limit expansions to expanding the coverage within LAION-5B (it is possible that many more EO images fulfilling the criteria existing in the 5 billion). Also, LAION-EO only contains images with at least 256 pixels in both width and height.

mikonvergence changed discussion status to closed

https://arxiv.org/abs/2306.11300 another big dataset (though this one contains few aerial imgs).
i'd be interested in training an equivalent of stable diffusion for remote sensing. would be cool to exchange a few ideas about it.
let me know if that would be something interesting for you, and we could have a chat about it.

Nice, I have been actively working on the same goal (among other projects) since StableDiffusion came out, but it is still work in progress.

Feel free to reach me on my professional e-mail mikolaj.czerkawski@esa.int if you want to chat!

Sign up or log in to comment