Instance segmentation is a major research topic in computer vision, therefore, having high-quality, large datasets is of paramount importance.

The most popular way of annotating instances is to use polygons for each instance.

Polygons can consist of a number of vertices ranging from just three to more than a hundred. They can have open spaces (holes) in them. Generally. their annotation is one of the most time-consuming aspects of the entire process. In fact, compared with another popular annotation task, such as object recognition with bounding boxes, the difference in the time spent annotating one instance is quite significant. It can (without any exaggerations) take 20 times more to annotate the given instance with a polygon compared to annotating it with a bounding box.

With this being said, a lot of research went into the automatic instance segmentation given the limited information about the object, e.g. its bounding box. It would be very time-efficient to simply annotate the bounding box and have a Neural Network (or another method) finish the annotation of the polygon and have the annotator fix the mistakes.

In Epoch 5, we are going to discuss such methods as well as how they can be implemented using the SuperAnnotate platform.