Segment Anything
One click, one mask.
In twenty twenty three, Meta released a model that could cut out almost any object in almost any image, from a single click. It didn't need to know what the object was. Drawing a good mask by hand takes minutes. With this, it took a second. For anyone who labels images, it changed the work overnight.
- 2023: Segment Anything
- A single click
- Minutes became seconds
How it works
The idea is a split. A heavy image encoder runs once, and produces a rich representation of the whole image. Then a very light decoder takes your click, and draws the mask, instantly. So a slow step happens once, and the quick step happens every time you click. That's why it feels so interactive.
- Heavy encoderruns once
- Your prompta click or a box
- Light decodermilliseconds
- Mask outseveral guesses
Using it in real work
SAM doesn't know what things are, only where their edges are. So teams combine it. A detector finds the apple, and SAM draws a perfect outline. Newer versions track objects through video too. And for your dataset, you click once per object, rather than drawing by hand. That one tool can cut labelling costs dramatically.
Summary
Separate the slow part from the fast part. Next, how to run a real vision project from start to finish.
Key takeaways- A heavy encoder runs once
- A light decoder per click
- Pair it with a detector