Image segmentation
이미지 세그멘테이션
Also known as: segmentation · instance segmentation
Separating the pixels that belong to each object in an image, answering where with an outline rather than a box.
There are two ways to handle objects in an image: detection, which marks position with a rectangle, and segmentation, which traces the region itself. The difference shows with overlapping or irregular shapes.
For a children's product the need is concrete. When a child taps the bed, only the bed should light up, and tapping one object in an illustration should not highlight its neighbour. A rectangle cannot do that.
Custom Cards and the picture book reader in NoryBook work this way: the region found for each object is lit on screen and the word card is attached to that spot.
- Versus detectionDetection gives a box, segmentation gives an outline.
- CostProducing masks costs more compute, so many systems enable it only where needed.