potato

https://github.com/celoyd/potato

Pansharpening is an image fusion task in satellite image processing, used on most of the satellite imagery on commercial maps. It’s illustrated below.

Potato is a pansharpening research project. It contains a working pansharpener, documentation, and training infrastructure. Potato is free to use and adapt for noncommercial purposes. It’s meant to be useful to people who need to process satellite imagery, but also interesting to others.

I love this bit:

Potato’s argument, in this analogy, is that a large fraction of existing pansharpening research and practice is trying to run in work boots. Potato aims to be useful as a pansharpener, but in the bigger picture its goal is to demonstrate a fresh way of thinking about pansharpening as a problem and to inspire research in that direction.

Specifically, Potato contends that pansharpening’s only audience is human perception, and therefore any accurate metric must be based distance in a perceptually uniform color space.

very relevant to LLMs as well, which measure themselves on benchmarks which are... questionably useful at best.