TL;DR
- I read this because.. : dataset filtering / evaluation์ ๋ํด ๊ถ๊ธํด์ ์ฝ์
- task : CLIP
- problem : open large image - text set
- idea : common crawl + study
- input/output : image / text -> similiarity score
- architecture : CLIP๊ณผ ๋์ผ
- objective : contrastive loss
- baseline : LAION-2B
- data : CommonPool 14B -> (filtered) DataComp 1.4B
- evaluation : zero-shot imagenet / imagenet-A/ .. ์๋ ์์ธํ ์์ + retrieval
- result : LAION-2B๋ณด๋ค ๋ ๋์ ์ฑ๋ฅ
- contribution : ๋ฐ์ดํฐ์ ๊ณต๊ฐ. ๋ค์ํ filtering ๊ธฐ๋ฒ ablation. competition์ผ๋ก ๋ฐ์ดํฐ์ ์ง์คํ๋ ์ฐ๊ตฌ ๋ฐฉํฅ ์ด์ง.
- etc. :
Details
Evaluation
- zs-image classifcation
- CLIP ์๋ ๋ ผ๋ฌธ์์ ํ๊ฐํ 22๊ฐ ๋ฐ์ดํฐ์
- 6๊ฐ์ distrbution shift๋ imagenet : ImaeNet-Sketch, ImageNet-V2, ImageNet-A, ImageNet-O, ImageNet-R, ObjectBet
- 13๊ฐ์ VTAB ๋ฐ์ดํฐ : https://arxiv.org/pdf/1910.04867.pdf
- 3๊ฐ์ WILDS ๋ฐ์ดํฐ: benchmark of 10 datasets reflecting a diverse range of distribution shifts that naturally arise in real-world applications, such as shifts across hospitals for tumor identification; across camera traps for wildlife monitoring; and across time and location in satellite imaging and poverty mapping. e.g. WILDS: A benchmark of in-the-wild distribution shifts. iWildCam2020-wilds(์ผ์๋๋ฌผ..), Camelyon17-wilds(์ธํฌ์กฐ์ง..), RxRx1-wilds(RNA…)
- WinoGAViL : commonsense association task https://paperswithcode.com/dataset/winogavil ๋ด๋ ๋ญ์ง ์ดํด๊ฐ ์๋๋น
- ๋ง์ง๋ง์ผ๋ก fairness ๋ฐ์ดํฐ ๋๊ฐ : FairFace, UTKFace -> ์ธ์ข ๋ง์ถ๋ classification
๋ช๊ฐ์ง ๋ฐ๊ฒฌ๋ค
zs retrieval๊ณผ linear probing์ ๋์ correlation
์์ ๋ฐ์ดํฐ์ ์ผ๋ก ํ ์ฑ๋ฅ๊ณผ ํฐ ๋ฐ์ดํฐ์ ์ผ๋ก ํ ์ฑ๋ฅ์ ๋์ correlation
imagenet๊ณผ ๋ค๋ฅธ ๋ฐ์ดํฐ์ ๊ฐ์ ๋์ correlation
correlation์ด ๋ฎ์ ์ ๋ค์ ์ฑ๋ฅ์ด random guess์ ๊ฐ๊น์ ๋ค.
- ๋ด๋ ์ดํด๊ฐ ์๋จ : https://paperswithcode.com/dataset/winogavil
- ์ผ์๋๋ฌผ : https://paperswithcode.com/dataset/iwildcam-2021
- ์์จ์ฃผํ : https://github.com/harshilpatel312/KITTI-distance-estimation
- ์ด๋ฏธ์ง๋ท์์ ์ค๋ถ๋ฅ๋ ๊ฑฐ ๋ชจ์๋์ ๊ฒ : https://paperswithcode.com/dataset/imagenet-a
- ์ธ๊ณต์์ฑ ์ด๋ฏธ์ง : https://paperswithcode.com/dataset/fmow
- ๋นํ๊ธฐ ์ข ๋ฅ ๋ถ๋ฅ : ttps://www.robots.ox.ac.uk/~vgg/data/fgvc-aircraft/
- ์ด ์ฌ์ง์ด ์ด๋ ๋๋ผ์์ ์ฐํ๋์ง ๋ถ๋ฅ : https://paperswithcode.com/dataset/country211
- ์๋ฃ์ชฝ : https://camelyon17.grand-challenge.org/ , https://patchcamelyon.grand-challenge.org/
- 3d ๋ฌผ๊ฑด๋ค ๊ด๊ณ : https://paperswithcode.com/dataset/clevr
๋ค ๋ํดํ๊ธฐ ์ง์ด ์๋ค.. ๊ทธ๋๋ง ์ฌ๊ธฐ์ ์ธ๋งํ๊ฑด imagenet-a์ country211 ์ ๋?! ๊ทธ๋ฆฌ๊ณ ๋น์ฐํ๊ฒ๋ ocr ์ชฝ ๋ฐ์ดํฐ์ (rendered sst2, svhn)๋ correlation์ด ใ ์์๋ค.
c.f. bs๊ณผ ๊ฐ์ hparam์ data filtering์ rank๋ ๊ฑฐ์ ๋ฐ๋์ง ์์๋ค