What is Annotation Grouping?

Annotation grouping is an effective way to label parts of the same object or instance that have been segregated due to truncation or occlusion by other objects in the scene. Without annotation grouping, each annotation instance will be considered separate, and a trained model would learn how to predict two halves of an object as two separate instances. With annotation grouping, however, the masks are merged and the model learns to predict both halves of an object as a singular instance. This can be useful in scenarios where the downstream task is object counting and the user requires accurate counts and minimize duplicate counts.

How to Group Annotations?

On the Nexus Annotator, simply Shift-Click multiple annotations of the same class. A white dotted border will appear that encompasses all selected annotations, and an option to Group (or Ungroup) the annotations will appear at the bottom of the canvas.

Once grouped, the annotations cannot be shifted around, edited, or deleted. If you wish to perform any of these operations on the annotations, simply click on the group and on the Ungroup button.