Semantic segmentation is a more generic term for the task of segmenting images into meaningful parts. It covers both object segmentation and instance segmentation, which were introduced in Chapter 1, Computer Vision and Neural Networks. Unlike image classification and object detection, covered in the previous chapters, segmentation tasks require the methods to return pixel-level dense predictions, that is, to assign a label to each pixel in the input images.
After explaining in more detail why encoders-decoders are thus great at object segmentation, and how their results can be further refined, we will present some solutions for the more complicated task of instance segmentation.
