Day 61: Semantic vs instance segmentation
Per-pixel understanding
Boxes are coarse; segmentation is per-pixel. Semantic segmentation labels every pixel with a class ('person', 'background', 'shirt') but doesn't separate two people. Instance segmentation additionally separates individual objects (person 1 vs person 2). For FitXpert's clean person cutout — needed for accurate silhouette and for Stage 4's try-on — you want a precise person mask, which segmentation provides where a box cannot.
Why the cutout matters twice
A precise person mask improves measurement (the silhouette edge is more accurate than a box) and is essential for Stage 4's virtual try-on (you composite the garment onto the exact person shape). This is the pipeline mindset paying dividends — a component built now serves a stage 70 days away.
Key terms
- Semantic segmentation
- Classifying every pixel into a category, without distinguishing separate object instances.
- Instance segmentation
- Segmentation that also separates individual object instances of the same class.
- Mask
- A per-pixel map marking which pixels belong to a given object or class.
What does instance segmentation provide that semantic segmentation does not?
