Semantic segmentation using transfer learning

In this recipe, we will take advantage of transfer learning and the fine-tuning of pretrained models to undertake a specific task of computer vision – the semantic segmentation of drone images.

Object detection and instance segmentation are about detecting objects in an image – an object is delimited by a bounding box, as well as a polygon in the case of instance segmentation. Alternatively, semantic segmentation is about classifying all the pixels of an image in a class.

As we can see in Figure 10.22, all pixels have a given color so that each one is attributed a class.

Figure 10.22 – An example of annotation of semantic segmentation. On the left is the original image, and on the right is the labeled image – there is one class of object per color, and each pixel is assigned to a given class

Even if it may look similar to instance segmentation, we will see in this recipe that...