Segmentation means labelling pixels rather than drawing boxes, but there are three different tasks under the name.
Semantic segmentation assigns every pixel a class from a fixed list and stops there.
It produces a map the same size as the image, coloured by category.
It has no notion of individual objects, so a crowd is one large region marked person.
That is enough for questions about area, such as how much of this field is planted.
Instance segmentation asks a different question: which pixels belong to this particular object.
Each detected object gets its own mask, so three people yield three masks and a count of three.
Most instance methods only handle countable things and simply ignore the rest of the image.
Vision research calls those countable categories things and the uncountable ones stuff.
Sky, road, grass and wall are stuff: they have no individual identity and cannot be counted.
Panoptic segmentation unifies the two by covering every pixel exactly once.
Things receive both a class and an instance identity, while stuff receives a class alone.
The result is a complete, non-overlapping description of the scene, which is what robotics tends to want.
The cost rises with ambition: per-pixel labelling is by far the most expensive annotation to collect.
Promptable models such as Segment Anything shifted much of that cost from labelling to prompting.