Segmentation basics

Segmentation labels pixels (or regions): what class each part of the image belongs to—finer than a single box.

Medical imaging assistance, photo editing, autonomous perception, and AR all need precise regions—plus careful validation.

Semantic segmentation: class per pixel (all “cars” same label). Instance segmentation: separate each object. Models often use encoder–decoder CNNs or transformer backbones.

Photo app “select subject” vs drawing one rectangle around it.

On a photo, outline two objects that share a class (two chairs). That’s why instances matter.

⚠️ Myth: Pixel-perfect masks mean clinical readiness.

✓ Reality: Clinical tools need regulatory validation far beyond demos.

⚠️ Myth: Segmentation always beats detection.

✓ Reality: Choose the cheapest representation that solves the job.

• Course 12 overview + detection; Course 05 CNNs
• Domain model cards (cite specifically)