These are pretty good results, but not SOTA currently EfficientNet-B7 outperforms this on ImageNet-1000. ImageNet in particular is around 84% top-1 and 97% top-

These are pretty good results, but not SOTA currently EfficientNet-B7 outperforms this on ImageNet-1000. ImageNet in particular is around 84% top-1 and 97% top-5. Claiming the last few percentages can be much harder than the first so it would be interesting to see this integrated into other approaches still. For comparison to other methods see, ImageNet Classification Leaderboards[1], Object Detection on COCO test-dev[2], Semantic Segmentation on COCO 2017[3]. Unfortunately COCO 2017 does not have many entries, I'm not aware of a better summary of SOTA methods. [1]: https://kobiso.github.io/Computer-Vision-Leaderboard/imagenet https://kobiso.github.io/Computer-Vision-Leaderboard/imagene... [2]: https://paperswithcode.com/sota/object-detection-on-coco https://paperswithcode.com/sota/object-detection-on-coco [3]: https://paperswithcode.com/sota/semantic-segmentation-on-coco-2017 https://paperswithcode.com/sota/semantic-segmentation-on-coc...