Semantic segmentation
The easiest way to think about segmentation is that it classifies each pixel in an image, assigning it to a corresponding class; combined, those pixels form areas of interest, such as regions with disease on an organ in medical images. By contrast, object detection (discussed in the previous section) classifies patches of an image into different object classes and creates bounding boxes around them.
We will demonstrate the modeling approach using data from the Sartorius – Cell Instance Segmentation competition (https://www.kaggle.com/c/sartorius-cell-instance-segmentation). In this one, the participants were tasked to train models for instance segmentation of neural cells using a set of microscopy images.
Our solution will be built around Detectron2, a library created by Facebook AI Research that supports multiple detection and segmentation algorithms.
Detectron2 is a successor to the original Detectron library (https://github.com/facebookresearch...
