Base-to-novel generalization on K-400
79.1Top-1 Acc (Base)TC-CLIP
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| TC-CLIPText Augmentation=LLM-based, Fine-tuning protocol=Directly fine-tuned from CLIP, Weight-space Ensembling=True2024.04 | 79.1 | 65.4 | 71.6 | |
| TC-CLIPText Augmentation=Original, Fine-tuning protocol=Directly fine-tuned from CLIP, Weight-space Ensembling=True2024.04 | 78.9 | 63.6 | 70.4 | |
| FROSTERText Augmentation=LLM-based, Fine-tuning protocol=Directly fine-tuned from CLIP, Weight-space Ensembling=True2024.04 | 77.8 | 64.3 | 70.4 | |
| Open-VCLIPText Augmentation=Original, Fine-tuning protocol=Directly fine-tuned from CLIP, Weight-space Ensembling=True2024.04 | 76.5 | 62.6 | 68.9 | |
| ViFi-CLIPText Augmentation=Original, Fine-tuning protocol=Directly fine-tuned from CLIP, Weight-space Ensembling=True2024.04 | 76.4 | 61.1 | 67.9 | |
| X-CLIPText Augmentation=Original, Fine-tuning protocol=Directly fine-tuned from CLIP, Weight-space Ensembling=True2024.04 | 74.1 | 56.4 | 64 | |
| A5Text Augmentation=Original, Fine-tuning protocol=Directly fine-tuned from CLIP, Weight-space Ensembling=True2024.04 | 69.7 | 37.6 | 48.8 | |
| Vanilla CLIPText Augmentation=Original, Fine-tuning protocol=Directly fine-tuned from CLIP, Weight-space Ensembling=True2024.04 | 62.3 | 53.4 | 57.5 | |
| ActionCLIPText Augmentation=Original, Fine-tuning protocol=Directly fine-tuned from CLIP, Weight-space Ensembling=True2024.04 | 61 | 46.2 | 52.6 |