Image Classification on VTAB (Granular Per-Dataset Accuracy)
89Caltech101 (All Acc)Oracle
Evaluation Results
| Method | Links | ||||||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| OracleBackbone=CLIP ViT-B/32, Evaluation Protocol=Fine-tuning, Training class availability=100% (All classes seen)2023.11 | 89 | 87.2 | 83.1 | 82.6 | 60 | 61.3 | 97.7 | 97.1 | 75.5 | 80.3 | 90.4 | 91.7 | 90.1 | 90.3 | 94.2 | 95.6 | 58.8 | 60 | 82.1 | 82.9 | |
| LOLSGD + LrankBackbone=CLIP ViT-B/32, Evaluation Protocol=Fine-tuning, Training class availability=50% missing2023.11 | 82.8 | 70.3 | 70 | 52.5 | 45.3 | 18.4 | 51.4 | 1.4 | 64.2 | 50.9 | 84.6 | 76.8 | 59.5 | 27.6 | 34.5 | 0.3 | 52.8 | 46.1 | 60.6 | 38.3 | |
| LOLSGDBackbone=CLIP ViT-B/32, Evaluation Protocol=Fine-tuning, Training class availability=50% missing2023.11 | 82.6 | 69.8 | 69.6 | 51.8 | 46 | 18.8 | 52.7 | 2.7 | 65.3 | 51 | 85.9 | 78.2 | 59.4 | 27.1 | 36.7 | 0.7 | 52.5 | 45.5 | 61.2 | 38.4 | |
| LOLSGD + Lrank + SEBackbone=CLIP ViT-B/32, Evaluation Protocol=Fine-tuning, Training class availability=50% missing2023.11 | 82.5 | 70.8 | 72.3 | 58.4 | 52.6 | 37.3 | 54.5 | 8.4 | 69.2 | 62.5 | 87.3 | 83.2 | 64.8 | 39.7 | 35.8 | 3.8 | 53.4 | 50 | 63.6 | 46 | |
| Naive TargetBackbone=CLIP ViT-B/32, Evaluation Protocol=Fine-tuning, Training class availability=50% missing2023.11 | 82.3 | 69.1 | 65.9 | 42.7 | 45.3 | 15.8 | 52.2 | 1.2 | 64.2 | 49.2 | 84.5 | 75.8 | 58.5 | 23.8 | 39.8 | 0.2 | 49.7 | 35.4 | 60.3 | 34.8 | |
| SGDBackbone=CLIP ViT-B/32, Evaluation Protocol=Fine-tuning, Training class availability=50% missing2023.11 | 82.3 | 69 | 66.1 | 42.7 | 45.5 | 16.4 | 52.5 | 1.8 | 64.5 | 49.8 | 84.7 | 75.8 | 57.9 | 21.9 | 40.1 | 0.5 | 49.8 | 36 | 60.4 | 34.9 | |
| LOLSGD + LdistillBackbone=CLIP ViT-B/32, Evaluation Protocol=Fine-tuning, Training class availability=50% missing2023.11 | 82 | 69.4 | 67.1 | 49.1 | 48 | 22.9 | 52.7 | 3.8 | 66.3 | 52.9 | 83.8 | 74.2 | 60 | 29 | 35.1 | 0.5 | 52.6 | 47.6 | 60.8 | 38.8 | |
| SourceBackbone=CLIP ViT-B/32, Evaluation Protocol=Zero-shot, Training class availability=50% missing2023.11 | 78.7 | 70.5 | 64.2 | 62.4 | 43.1 | 47 | 32.2 | 20.3 | 63.8 | 68.4 | 84.1 | 88.7 | 54 | 52.7 | 8.8 | 11.3 | 46.7 | 49.1 | 52.8 | 52.3 |