Polygons are used in the models where the objects do not have a regular shape, and require precision.
Challenges and solutions:
* Pixel-level accuracy. Our custom designed image annotation tools allow drawing precisely the borders of the object and record the maximim x/y points for the Deep Learning model. COCO formatallows machine learning engineers to use the appropriate shape in class training, interchangeably between using the output recorded as rectangle or multi-point polygon.
* Geospatial. If the the polygonal annotation is a part of physical footprint mapping, where the borders of the objects (container ships, cranes) matter, choosing polygon annotation as a tool is important. In addition, to free line polygon, at TaQadam we used ellipses and other standard shapes to increase the speed.