Multi-label Classification on ImageNet ML v2
88.2Top-1 Accuracy+Multi-label FT
Evaluation results
| Method | Links | ||
|---|---|---|---|
| +Multi-label FTModel=ViT-large2026.03 | 88.2 | 83.3 | |
| Multi-label E2EModel=ViT-large2026.03 | 87.9 | 83 | |
| Single-label E2EModel=ViT-large2026.03 | 87.1 | 81.4 | |
| +Multi-label FTModel=ViT-base2026.03 | 86.9 | 81.3 | |
| +IN1k-Mul FTBackbone=ViT-B/16, Resolution=224, Pre-training=ImageNet-21K (MIIL), Supervision=Multi-label (Mul), Training Protocol=Fine-tuning2026.03 | 86.9 | 82.6 | |
| Multi-label E2EModel=ViT-base2026.03 | 86.8 | 81.8 | |
| IN1k-Mul E2EBackbone=ViT-B/16, Resolution=224, Supervision=Multi-label (Mul), Training Protocol=End-to-end2026.03 | 86.8 | 81.8 | |
| +IN1k-Sig FTBackbone=ViT-B/16, Resolution=224, Pre-training=ImageNet-21K (MIIL), Supervision=Single-label (Sig), Training Protocol=Fine-tuning2026.03 | 86.4 | 80.7 | |
| Single-label E2EModel=ViT-base2026.03 | 85.8 | 80.3 | |
| Multi-label E2EModel=ViT-small2026.03 | 85.1 | 80.7 | |
| +Multi-label FTModel=ViT-small2026.03 | 84.1 | 80.2 | |
| Single-label E2EModel=ViT-small2026.03 | 83.1 | 75.6 | |
| +Multi-label FTModel=ResNet-1012026.03 | 82.4 | 77.7 | |
| Multi-label E2EModel=ResNet-1012026.03 | 82.3 | 77.4 | |
| Multi-label E2EModel=ResNet-502026.03 | 81 | 76.2 | |
| Multi-labelBackbone=ResNet-502026.03 | 81 | 76.2 | |
| +Multi-label FTModel=ResNet-502026.03 | 80.4 | 76.1 | |
| Single-label E2EModel=ResNet-1012026.03 | 80 | 72.7 | |
| ReLabel w/ Our MaskBackbone=ResNet-502026.03 | 79.7 | 74.6 | |
| ReLabelBackbone=ResNet-502026.03 | 79.4 | 74.8 | |
| Single-label E2EModel=ResNet-502026.03 | 78.2 | 72.3 | |
| Original + Label SmoothBackbone=ResNet-502026.03 | 78.2 | 72.3 | |
| LLBackbone=ResNet-502026.03 | 77.7 | 72.7 | |
| Original LabelBackbone=ResNet-502026.03 | 77.4 | 73 |