Skip to content
Computer Vision

Seeing More Than Labels

Segmentation: Pixel-Perfect Masks

9 min+85 XPStep 1/10

A classifier for every pixel

Segmentation asks the model to label every pixel. Two flavors matter:

  • Semantic segmentation — label pixels by category: road, sidewalk, car, sky. Two cars touching are both just 'car pixels'.
  • Instance segmentation — also separate the individuals: car #1 and car #2 get different masks, even where they overlap.

A surgeon doesn't want to know a scan 'contains a tumor' — they need its exact boundary. A self-driving car doesn't want 'there's road somewhere' — it needs the precise outline of the drivable surface.