concept semantic segmentation in category machine learning
This is an excerpt from Manning's book Human-in-the-Loop Machine Learning MEAP V09.
Semantic segmentation is when the entire image receives a label, with accurate polygons boundaries around all objects. Because it labels every pixel in the image, it is also referred to as “pixel labeling”. An example is given in Figure 6.7.
If you are trying to estimate objects that extend behind some other object (occlusion) then it is more common to use the bounding-box type object detection you learned about in the last section. It is also more common to paint all objects as a single type with semantic segmentation rather than separately identify each object. For example, every tree in Figure 6.6 is the same color, but the image is not distinguishing one tree from another. However, there are cases where bounding boxes are used that ignore occlusion, where semantic segmentation does try to capture occlusion, and also where semantic segmentation does distinguish objects (called “instance segmentation”). If a model combines all of these methods it is sometimes known as “panoptic segmentation”, identifying objects and background pixels.
All the methods in this chapter should be generic enough that they can apply to any variation in how bounding boxes or semantic segmentation is used.
Figure 6.7: An example of semantic segmentation, where every pixel is labeled (hence also known as “pixel labeling”). This kind of colored photograph is what a lot of semantic segmentation tools look like: a coloring-in exercise. We’ll cover those tools later in the chapters on annotation. If you’re looking at this image in black-and-white, the contrastive shades of grey should give you a good enough idea of what it would look like in color. If the different objects receive a label, for example if the four trees are labeled separately, the task is known as Instance Segmentation.
