Prodigy-Segment tutorial: pixel masks and caching

This tutorial, uploaded on December 14, 2023, demonstrates Prodigy-Segment, a Prodigy plugin that uses Meta's Segment Anything model for pixel annotation. No software version or separate recording date is supplied, so the interface and hardware guidance describe the demonstrated setup at that time.

The presenter starts a recipe with a dataset name, an image folder, model weights and annotation labels. In the interface, they draw selections around a cat and a dog, then press the segment button to generate pixel masks. They show that selections can be adjusted or removed before accepting the example. Loading the saved dataset reveals separate segmentation maps for the two animals, including a small area of dog fur behind the cat's ear.

The final section covers the hardware tradeoff for this local AI workflow. The speaker says real-time segmentation may require a GPU because the model is large. A fill-cache recipe performs precomputation and stores the results on disk for use during annotation. The presenter recommends running it ahead of time to reduce hardware demands and keep annotation responsive. They say the recipes should work on modern laptops without a GPU, but still require substantial memory; the tutorial gives no numerical RAM requirement.