Image Classification on ImageNet 1k (val) (Linear & k-NN Probing)
82.2Linear AccuracyMAE-CT_aug
Evaluation results
| Method | Links | ||
|---|---|---|---|
| MAE-CT_augArchitecture=ViT-H/16, Epochs=800; 40, GV/LV/T=2, Mask ratio=0.75; 0, Image Augmentation=extensive2023.04 | 82.2 | 79.8 | |
| MAE-CT_minArchitecture=ViT-H/16, Epochs=800; 30, GV/LV/T=2, Mask ratio=0.75; 0, Image Augmentation=minimal2023.04 | 81.5 | 79.4 | |
| MAE-CT_augArchitecture=ViT-L/16, Epochs=800; 40, GV/LV/T=2, Mask ratio=0.75; 0, Image Augmentation=extensive2023.04 | 81.5 | 79.1 | |
| SupervisedArchitecture=Swin-T, Parameters=28M, epoch=300, Throughput (im/s)=7552021.06 | 81.2 | — | |
| I-JEPAArchitecture=ViT-H/16_448, Epochs=300, GV/LV/T=1*/4*/1, Mask ratio=*, Image Augmentation=minimal2023.04 | 81.1 | 70.1 | |
| MoCo v3Architecture=ViT-L/7-BN, Epochs=300, GV/LV/T=2, Image Augmentation=extensive2023.04 | 81 | — | |
| iBOTArchitecture=ViT-L/16, Epochs=250, GV/LV/T=2/10/2, Mask ratio=≈ 0.15†, Image Augmentation=extensive2023.04 | 81 | 78 | |
| Layer GraftingArchitecture=ViT-L/16, Epochs=1600; 300, GV/LV/T=1;2/-/-;2, Mask ratio=0.75; 0, Image Augmentation=extensive2023.04 | 81 | 77.3 | |
| DINO + XCiT-M24/8 | 384SSL Method=DINO, Backbone=XCiT-M24/8, #params=84M, FLOPS=188.0B, Input Resolution=384x384, Crop-ratio=1.0, Training epochs=3002021.06 | 80.9 | — | |
| MSNArchitecture=ViT-L/7, Epochs=200, GV/LV/T=1/10/1, Mask ratio=0.7, Image Augmentation=extensive2023.04 | 80.7 | — | |
| DINO + XCiT-M24/8SSL Method=DINO, Backbone=XCiT-M24/8, #params=84M, FLOPS=64.0B, Input Resolution=224x224, Crop-ratio=0.875, Training epochs=3002021.06 | 80.3 | 77.9 | |
| MAE-CT_minArchitecture=ViT-L/16, Epochs=800; 20, GV/LV/T=2, Mask ratio=0.75; 0, Image Augmentation=minimal2023.04 | 80.2 | 78 | |
| DINO + ViT-B/8SSL Method=DINO, Backbone=ViT-B/8, #params=87M, FLOPS=78.2B, Input Resolution=224x224, Crop-ratio=0.875, Training epochs=3002021.06 | 80.1 | 77.4 | |
| DINOArchitecture=ViT-B/8, Epochs=400, GV/LV/T=2/10/2, Image Augmentation=extensive2023.04 | 80.1 | 77.4 | |
| iBOTArchitecture=ViT-B/16, Epochs=400, GV/LV/T=2/10/2, Mask ratio=≈ 0.15†, Image Augmentation=extensive2023.04 | 79.5 | 77.1 | |
| I-JEPAArchitecture=ViT-H/14, Epochs=300, GV/LV/T=1*/4*/1, Mask ratio=*, Image Augmentation=minimal2023.04 | 79.3 | 71.6 | |
| DINO + ViT-S/8SSL Method=DINO, Backbone=ViT-S/8, #params=22M, FLOPS=22.4B, Input Resolution=224x224, Crop-ratio=0.875, Training epochs=3002021.06 | 79.2 | 77.2 | |
| DINO + XCiT-S12/8SSL Method=DINO, Backbone=XCiT-S12/8, #params=26M, FLOPS=18.9B, Input Resolution=224x224, Crop-ratio=0.875, Training epochs=3002021.06 | 79.2 | 77.1 | |
| MoCo v3Architecture=ViT-H/14-BN, Epochs=300, GV/LV/T=2, Image Augmentation=extensive2023.04 | 79.1 | — | |
| DINO + XCiT-M24/16SSL Method=DINO, Backbone=XCiT-M24/16, #params=84M, FLOPS=16.2B, Input Resolution=224x224, Crop-ratio=0.875, Training epochs=3002021.06 | 78.8 | 76.4 | |
| DINOArchitecture=ViT-S/8, Epochs=400, GV/LV/T=2/10/2, Mask ratio=79.7, Image Augmentation=extensive2023.04 | 78.3 | — | |
| DINO + ViT-B/16SSL Method=DINO, Backbone=ViT-B/16, #params=87M, FLOPS=17.5B, Input Resolution=224x224, Crop-ratio=0.875, Training epochs=3002021.06 | 78.2 | 76.1 | |
| MoCo v3Architecture=ViT-H/14, Epochs=300, GV/LV/T=2, Image Augmentation=extensive2023.04 | 78.1 | — | |
| MAE (reimpl.)Architecture=ViT-H/16, Epochs=800, GV/LV/T=2, Mask ratio=0.75, Image Augmentation=minimal2023.04 | 78 | 61.1 | |
| Layer GraftingArchitecture=ViT-B/16, Epochs=1600; 300, GV/LV/T=1;2/-/-;2, Mask ratio=0.75; 0, Image Augmentation=extensive2023.04 | 77.9 | 75.4 | |
| DINO + XCiT-S12/16SSL Method=DINO, Backbone=XCiT-S12/16, #params=26M, FLOPS=4.9B, Input Resolution=224x224, Crop-ratio=0.875, Training epochs=3002021.06 | 77.8 | 76 | |
| MoCo v3Architecture=ViT-L/16, Epochs=300, GV/LV/T=2, Image Augmentation=extensive2023.04 | 77.6 | — | |
| I-JEPAArchitecture=ViT-L/16, Epochs=600, GV/LV/T=1*/4*/1, Mask ratio=*, Image Augmentation=minimal2023.04 | 77.5 | — | |
| MAEArchitecture=ViT-H/14, Epochs=1600, GV/LV/T=1, Mask ratio=0.75, Image Augmentation=minimal2023.04 | 77.2 | 58.9 | |
| DINOArchitecture=DeiT-S, Parameters=21M, epoch=800, Throughput (im/s)=1007, Evaluation Protocol=linear probing of DINO2021.06 | 77 | 74.5 | |
| MSTArchitecture=DeiT-S, Parameters=21M, epoch=300, Throughput (im/s)=1007, Evaluation Protocol=linear probing of DINO2021.06 | 76.9 | 75 | |
| MAE-CT_augArchitecture=ViT-B/16, Epochs=800; 80, GV/LV/T=2, Mask ratio=0.75; 0, Image Augmentation=extensive2023.04 | 76.9 | 73.4 | |
| MoCo v3Architecture=ViT-B/16, Epochs=600, GV/LV/T=2, Image Augmentation=extensive2023.04 | 76.7 | 72.6 | |
| SupervisedArchitecture=ResNet-50, Parameters=23M, epoch=100, Throughput (im/s)=12372021.06 | 76.5 | — | |
| NNCLRArchitecture=ViT-B/16, Epochs=1000, GV/LV/T=2, Image Augmentation=extensive2023.04 | 76.5 | — | |
| SupervisedArchitecture=DeiT-S, Parameters=21M, epoch=300, Throughput (im/s)=10072021.06 | 76.4 | — | |
| MSTArchitecture=DeiT-S, Parameters=21M, epoch=300, Throughput (im/s)=10072021.06 | 76.3 | 75 | |
| DINO + ViT-S/16SSL Method=DINO, Backbone=ViT-S/16, #params=22M, FLOPS=4.6B, Input Resolution=224x224, Crop-ratio=0.875, Training epochs=3002021.06 | 76.1 | 72.8 | |
| DINOArchitecture=ViT-B/16, Epochs=400, GV/LV/T=2/10/2, Mask ratio=78.2, Image Augmentation=extensive2023.04 | 76.1 | — | |
| MAEArchitecture=ViT-L/16, Epochs=1600, GV/LV/T=1, Mask ratio=0.75, Image Augmentation=minimal2023.04 | 76 | 60.6 | |
| DINOArchitecture=DeiT-S, Parameters=21M, epoch=300, Throughput (im/s)=1007, Evaluation Protocol=linear probing of DINO2021.06 | 75.9 | 72.8 | |
| MAE (reimpl.)Architecture=ViT-L/16, Epochs=800, GV/LV/T=2, Mask ratio=0.75, Image Augmentation=minimal2023.04 | 75.9 | 57.6 | |
| SwAVArchitecture=ResNet-50, Parameters=23M, epoch=800, Throughput (im/s)=12372021.06 | 75.3 | 65.7 | |
| DINOArchitecture=DeiT-S, Parameters=21M, epoch=300, Throughput (im/s)=10072021.06 | 75.2 | 72.8 | |
| MoBY + Swin-TSSL Method=MoBY, Backbone=Swin-T, #params=29M, FLOPS=4.5B, Input Resolution=224x224, Crop-ratio=0.875, Training epochs=3002021.06 | 75 | — | |
| MSTArchitecture=DeiT-S, Parameters=21M, epoch=100, Throughput (im/s)=1007, Evaluation Protocol=linear probing of DINO2021.06 | 75 | 72.1 | |
| DINO + ResNet-50SSL Method=DINO, Backbone=ResNet-50, #params=23M, FLOPS=4.1B, Input Resolution=224x224, Crop-ratio=0.875, Training epochs=3002021.06 | 74.5 | 65.6 | |
| BYOLArchitecture=ResNet-50, Parameters=23M, epoch=1000, Throughput (im/s)=12372021.06 | 74.4 | 64.8 | |
| CMAEArchitecture=ViT-B/16, Epochs=1600, GV/LV/T=1/-/1, Mask ratio=0.75, Image Augmentation=extensive2023.04 | 73.9 | — | |
| MSTArchitecture=Swin-T, Parameters=28M, epoch=100, Throughput (im/s)=7552021.06 | 73.8 | 66.2 | |
| SwAVArchitecture=DeiT-S, Parameters=21M, epoch=800, Throughput (im/s)=10072021.06 | 73.5 | 66.3 | |
| MAE-CT_minArchitecture=ViT-B/16, Epochs=800; 20, GV/LV/T=2, Mask ratio=0.75; 0, Image Augmentation=minimal2023.04 | 73.5 | 64.1 | |
| I-JEPAArchitecture=ViT-B/16, Epochs=600, GV/LV/T=1*/4*/1, Mask ratio=*, Image Augmentation=minimal2023.04 | 72.9 | — | |
| MOBYArchitecture=DeiT-S, Parameters=21M, epoch=300, Throughput (im/s)=10072021.06 | 72.8 | — | |
| MoCov2Architecture=DeiT-S, Parameters=21M, epoch=800, Throughput (im/s)=10072021.06 | 72.7 | 64.4 | |
| MoCov3Architecture=DeiT-S, Parameters=21M, epoch=300, Throughput (im/s)=10072021.06 | 72.5 | — | |
| BYOLArchitecture=DeiT-S, Parameters=21M, epoch=800, Throughput (im/s)=10072021.06 | 71.4 | 66.6 | |
| MoCov2Architecture=ResNet-50, Parameters=23M, epoch=800, Throughput (im/s)=12372021.06 | 71.1 | 61.9 | |
| BYOLArchitecture=DeiT-S, Parameters=21M, epoch=300, Throughput (im/s)=10072021.06 | 71 | — | |
| MOBYArchitecture=Swin-T, Parameters=28M, epoch=100, Throughput (im/s)=7552021.06 | 70.9 | 57.34 | |
| SimCLRArchitecture=DeiT-S, Parameters=21M, epoch=300, Throughput (im/s)=10072021.06 | 69 | — | |
| MAEArchitecture=ViT-B/16, Epochs=1600, GV/LV/T=1, Mask ratio=0.75, Image Augmentation=minimal2023.04 | 68 | 50 | |
| SwAVArchitecture=DeiT-S, Parameters=21M, epoch=300, Throughput (im/s)=10072021.06 | 67.1 | — | |
| MAE (reimpl.)Architecture=ViT-B/16, Epochs=800, GV/LV/T=2, Mask ratio=0.75, Image Augmentation=minimal2023.04 | 66.7 | 51.1 | |
| SimMIMArchitecture=ViT-B/16, Epochs=800, GV/LV/T=1, Mask ratio=0.5, Image Augmentation=minimal2023.04 | 56.7 | — |