Vision Models Are More Robust And Fair When Pretrained On Uncurated Images Without Supervision

Open paper
Benchmark Model Rank Results
action-classification-on-kinetics-700SEER (RegNet10B)#28Top-1 Accuracy: 51.9
domain-generalization-on-imagenet-aSEER (RegNet10B)#17Top-1 accuracy %: 52.7
domain-generalization-on-imagenet-rSEER (RegNet10B)#18Top-1 Error Rate: 43.9
domain-generalization-on-imagenet-sketchSEER (RegNet10B)#14Top-1 accuracy: 45.6
fine-grained-image-classification-on-caltechSEER (RegNet10B - linear eval)#10Top-1 Error Rate: 9.0%Accuracy: 91.0
fine-grained-image-classification-on-fgvcSEER (RegNet10B)#40Accuracy: 54.82%
fine-grained-image-classification-on-oxford-1SEER (RegNet10B)#12Accuracy: 85.3%
fine-grained-image-classification-on-stanfordSEER (RegNet10B)#66Accuracy: 68.03%
fine-grained-image-classification-on-sun397SEER (RegNet10B - linear eval)#2Accuracy: 80.0
image-classification-on-cifar-10SEER (RegNet10B)#191Percentage correct: 90
image-classification-on-cifar-100SEER (RegNet10B)#112Percentage correct: 81.53
image-classification-on-dtdSEER (RegNet10B - linear eval)#6Accuracy: 80.5
image-classification-on-eurosatSEER (RegNet10B - linear eval)#13Accuracy (%): 97.5
image-classification-on-flowers-102SEER (RegNet10B)#40Accuracy: 96.3
image-classification-on-food-101-1SEER (RegNet10B - linear eval)#2Accuracy (%): 90.3
image-classification-on-imagenetSEER (RG-10B)#188Top 1 Accuracy: 85.8%Number of params: 10000M
image-classification-on-imagenet-realSEER (RegNet10B)#22Accuracy: 89.8%Params: 10000M
image-classification-on-imagenet-v2SEER (RegNet10B)#17Top 1 Accuracy: 76.2
image-classification-on-inaturalist-2018SEER (RegNet10B - finetuned - 384px)#7Top-1 Accuracy: 84.7%
image-classification-on-mnistSEER (RegNet10B)#30Percentage error: 0.58Accuracy: 99.42
image-classification-on-objectnetSEER (RegNet10B)#17Top-1 Accuracy: 60.2
image-classification-on-places205SEER (RegNet10B - finetuned - 384px)#3Top 1 Accuracy: 69.0
image-classification-on-resisc45SEER (RegNet10B)#5Top 1 Accuracy: 95.61
image-classification-on-resisc45SwAV (ResNet50-w5)#8Top 1 Accuracy: 94.73
image-classification-on-resisc45DINO (DeiT-B/16)#9Top 1 Accuracy: 93.97
image-classification-on-resisc45MoCo-v3 (ViT-B/16)#11Top 1 Accuracy: 93.35
image-classification-on-resisc45CLIP (ViT-B/16)#12Top 1 Accuracy: 92.7
image-classification-on-resisc45DeiT-B/16#13Top 1 Accuracy: 92.48
image-classification-on-resisc45SimCLR-v2 (ResNet152-w3 + SK)#14Top 1 Accuracy: 89.77
image-classification-on-resisc45ResNet50 (ImageNet-supervised)#15Top 1 Accuracy: 88.56
image-classification-on-resisc45MoCo-v2 (ResNet50)#17Top 1 Accuracy: 85.4
image-classification-on-stl-10SEER (RegNet10B)#5Percentage correct: 97.3PARAMS: 10000M
image-classification-on-svhnSEER (RegNet10B)#39Percentage error: 13.6
meme-classification-on-hateful-memesSEER (RegNet10B)#14ROC-AUC: 0.734
self-supervised-image-classification-on-1SEER (Regnet10B)#20Top 1 Accuracy: 85.8%Number of Params: 10000M
self-supervised-image-classification-on-imagenetSEERv2#27Top 1 Accuracy: 79.8%Number of Params: 10000M
semi-supervised-image-classification-on-1SEER (RegNet10B)#30Top 1 Accuracy: 62.4%
semi-supervised-image-classification-on-2SEER (RegNet10B)#12Top 1 Accuracy: 78.8%
traffic-sign-recognition-on-gtsrbSEER (RegNet10B)#5Accuracy: 90.71%

Rank counts only results with a code link.