Optimizing scales vs. overlap

[JohnMBrandt]

Great tool! I work on small object detection for highly variably sized objects, and the anchor sizes have a lot of influence on the output.

I've noticed that, at least for my work, getting consistently high overlap, as generated through the train-config evaluation function, is much more important than the optimization done in pyodi train-config generation.

More specifically, the outputs of generation fit the bounding box size distribution of the data, but have such low overlap that the boxes are never assigned to a positive class. Manually playing around with the scales and ratios and regenerating the overlap figures until the whole dataset had consistently high overlap resulted in quite a bit of improved accuracy. Curious if this is something you've discussed optimizing for?