Deep Image Clustering on ImageNet 50 (val)
98.72NMIProbing CLIP ViT-L/14
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| Probing CLIP ViT-L/14Backbone=ViT-L/14, Pre-training strategy=Natural language supervised (CLIP), Protocol=Linear Probing2023.03 | 98.72 | 98.96 | 97.88 | |
| Probing Sup. ViT-L/16Backbone=ViT-L/16, Pre-training strategy=Supervised, Protocol=Linear Probing2023.03 | 97.77 | 98.12 | 96.21 | |
| TEMI Sup. ViT-L/16Backbone=ViT-L/16, Pre-training strategy=Supervised, Protocol=Clustering2023.03 | 95.75 | 95.12 | 91.4 | |
| Probing DINO VIT-B/16Backbone=ViT-B/16, Pre-training strategy=Self-supervised (DINO), Protocol=Linear Probing2023.03 | 95.1 | 95.76 | 91.64 | |
| Probing MSN ViT-L/16Backbone=ViT-L/16, Pre-training strategy=Self-supervised (MSN), Protocol=Linear Probing2023.03 | 94.21 | 94.92 | 90.03 | |
| TEMI CLIP ViT-L/14Backbone=ViT-L/14, Pre-training strategy=Natural language supervised (CLIP), Protocol=Clustering2023.03 | 92.32 | 88.27 | 82.78 | |
| TEMI MSN VIT-L/16Backbone=ViT-L/16, Pre-training strategy=Self-supervised (MSN), Protocol=Clustering2023.03 | 88.14 | 84.87 | 76.46 | |
| TEMI DINO VIT-B/16Backbone=ViT-B/16, Pre-training strategy=Self-supervised (DINO), Protocol=Clustering2023.03 | 86.1 | 80.01 | 70.93 | |
| ProposBackbone=ResNet-50, Protocol=Clustering2023.03 | 82.8 | — | 69.1 | |
| SCANBackbone=ResNet-50, Protocol=Clustering2023.03 | 82.2 | 76.8 | 66.1 |