Skip to main content

Supersize Me: Bridging the Train-Test Augmentation Gap

·76 words·1 min

A nice paper on simple observation that augs for train and test are different causing distributional shift. They propose simple trick: just increasing test time image size. If you are also willing to do fine tuning on that size, gains become significant!

https://arxiv.org/abs/1906.06423

A follow up of this paper now holds the new ImageNet state of the art at 88.5% top1 and 98.7% top5 by applying this method on previous state of the art.

https://arxiv.org/abs/2003.08237

Discussion