Most of our anchor boxes are 28x28, we should try with smaller anchor boxes to see if we can get an improvement out of our network somehow.
We could also consider not using certain levels of our Feature Pyramid Network as I think lower(??) levels correspond to detecting larger areas of the image.
Most of our anchor boxes are 28x28, we should try with smaller anchor boxes to see if we can get an improvement out of our network somehow.
We could also consider not using certain levels of our Feature Pyramid Network as I think lower(??) levels correspond to detecting larger areas of the image.