Chapter 4 · The deep learning era (2006–2016)
Image segmentation
Segmentation labels every pixel, outlining the exact shape of each object.
Chapter 4 · The deep learning era (2006–2016)
Segmentation labels every pixel, outlining the exact shape of each object.
A photo is just pixels…
Detection draws a rectangle. Segmentation goes further and outlines the true shape, pixel by pixel.
A box around a curved object also holds a lot of background. For many jobs, that is not precise enough.
Tap or drag across the picture (a finger is easier than a single square). The mask spreads to neighbours that look alike. The slider sets how alike they must be.
A clean mask. Compare it with the box. Simplified teaching version: real models learn shapes, they do not just compare brightness.
What does segmentation label?
Why can a box be a poor fit for a curved object?
Two cats stand side by side. Which kind separates them as cat 1 and cat 2?