Great work, great article.

My question is how would you imagine this being used at a more complex problem, such as object detection / segmentation. Large unlabeled datasets might be available, but you still need some kind of label to include them in your dataset.

Also, have you compared how this approach performs if you lift the 10k cap to, say, 20k? How well does the model perform if you go all the way and add all 1M augmentations? My point is, how close is the 10k-capped performance to what you would get with no dataset size restriction