Minimum/maximum object bounding box size


Can anyone provide any general guidance on how small of an object
should be labeled? Should the “size” parameter be a function of
percentage of square pixels to the whole image size? Would a too
small of a labeled object be detrimental to the model’s performance if
it is too small? I envision if an object is too small, there is not
sufficient detail in the boxed object to properly train the model and
be counter productive. Is this correct?

Additionally, is it detrimental to the model’s performance if the
bounding box for an isolated labeled object is larger than the object
itself? When is over sized too large?


Hi, we have blog posts about these two topics here:

I’ll add that you’ll want to localize your bounding boxes to the objects as much as possible, so “large” is anything in which you’re capturing extra information for no reason at all with your bounding boxes. Tight boxes are best!

The model’s performance will be negatively impacted if bounding boxes are drawn with little care. Improving Computer Vision Datasets and Models