I was wondering if the training data gets (significantly) better if the objects are annotated with a perfect outline vs using a square box?
Hi @Birdo - thanks for posting!
If you are training an object detection model, we will take the polygon annotation and fit the tightest possible bounding box to it. This will be significantly more accurate in cases where you use rotation augmentations (if you just give a box, we rotate the box and then re-fit a new bounding box).
If you are training an instance segmentation model, you need the polygon annotations.
This topic was automatically closed 21 days after the last reply. New replies are no longer allowed.