Base-to-New Classification on 11 Downstream Datasets (Imbalanced, τ=0.25)
72.62Accuracy (IN.)CPT
Evaluation results
| Method | Links | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| CPTBackbone=ViT-B/16, Training protocol=trained from base classes, Max samples per class (max{nk})=16, Evaluation Metric=Harmonic Mean (H)2026.05 | 72.62 | 96.44 | 96.16 | 71.58 | 79.37 | 91.13 | 30.16 | 77.65 | 57.74 | 59.09 | 77.42 | 73.58 | |
| NPTBackbone=ViT-B/16, Training protocol=trained from base classes, Max samples per class (max{nk})=16, Evaluation Metric=Harmonic Mean (H)2026.05 | 72.22 | 95.66 | 96.2 | 70.89 | 78.58 | 90.23 | 29.86 | 76.88 | 57.16 | 58.48 | 76.65 | 72.98 | |
| DeKgBackbone=ViT-B/16, Training protocol=trained from base classes, Max samples per class (max{nk})=16, Evaluation Metric=Harmonic Mean (H)2026.05 | 72.08 | 95.86 | 95.3 | 71.12 | 78.98 | 90.46 | 29.91 | 76.9 | 57.58 | 58.61 | 76.11 | 72.99 | |
| DPCBackbone=ViT-B/16, Training protocol=trained from base classes, Max samples per class (max{nk})=16, Evaluation Metric=Harmonic Mean (H)2026.05 | 71.85 | 95.35 | 96.26 | 71.41 | 78.11 | 90 | 29.74 | 77.18 | 57.08 | 58.85 | 76.28 | 72.92 | |
| CoPromptBackbone=ViT-B/16, Training protocol=trained from base classes, Max samples per class (max{nk})=16, Evaluation Metric=Harmonic Mean (H)2026.05 | 71.5 | 94.71 | 95.24 | 70.2 | 77.77 | 89.35 | 29.56 | 76.12 | 56.58 | 57.89 | 75.69 | 72.24 | |
| MaPLeBackbone=ViT-B/16, Training protocol=trained from base classes, Max samples per class (max{nk})=16, Evaluation Metric=Harmonic Mean (H)2026.05 | 70.93 | 95.38 | 94.22 | 70.25 | 76.59 | 88.71 | 28.62 | 75.98 | 55.15 | 53.97 | 75.43 | 71.38 |