I would like to see a "Magic Wand" tool for imagery that uses "few-shot" learning. This would allow users to identify features in satellite or drone imagery without needing hundreds of training samples.
Use Case / Workflow:
Current Workflow: To find swimming pools in a city, I have to digitize 200+ training samples, run a training tool for hours, and then run an inference tool.
Proposed Workflow: I click on 3 pools. The AI scans the current map extent and highlights all similar pixels. I click "Reject" on a blue rooftop it mistakenly identified, and the model refines itself instantly.
