What is Instance Segmentation?
Instance segmentation assigns pixels both a semantic class and a distinct object identity. Unlike semantic segmentation, it separates multiple objects of the same class into individual masks.
How Instance Segmentation works
An instance-aware model produces a separate spatial mask for each detected object together with its category and confidence. Architectures may first propose objects and predict masks per proposal, or derive identities from a unified pixel representation. This output preserves count and overlap relationships that a class-only region map loses. It feeds object editing, inventory, robotics, tracking, and measurement after inference and mask postprocessing.
Key facts
- Panoptic segmentation combines countable object instances with non-instance background regions, while instance segmentation can omit amorphous background classes entirely.
- Two partially occluded objects of one category require separate identity hypotheses; a model may merge them into one mask even when its combined class area looks correct.
- Instance identifiers are categorical, so mask resizing or export must preserve exact IDs; interpolating the identifier image creates invalid intermediate object numbers.
When Instance Segmentation matters
Use instance segmentation when a system must count, edit, measure, or track separate objects rather than one combined class region. Overlapping or partially hidden objects can be merged, split, or missed, affecting downstream totals and masks.