Image-to-Text Retrieval on ICR
62.5R@1Chinese CLIP
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| Chinese CLIPBackbone=ViT-L/14, Evaluation Protocol=Finetuned2022.11 | 62.5 | 83.9 | 89.1 | 78.3 | |
| R2D2Backbone=ViT-L/14, Evaluation Protocol=Finetuned2022.11 | 61.5 | 82.9 | 87.7 | 77 | |
| R2D2Backbone=ViT-L, Zero-shot=true2022.05 | 58 | 80.5 | 85.2 | — | |
| Chinese CLIPBackbone=ViT-B/16, Evaluation Protocol=Finetuned2022.11 | 56.6 | 79.5 | 85.6 | 73.5 | |
| R2D2Backbone=ViT-B, Evaluation Protocol=Finetuned2022.11 | 43.4 | 69.8 | 78.4 | 63.5 | |
| WukongBackbone=ViT-L, Zero-shot=true2022.05 | 35.1 | 58.2 | 66.3 | — | |
| CLIPBackbone=ViT-L, Zero-shot=true2022.05 | 30.3 | 52.9 | 61.6 | — | |
| FILIPBackbone=ViT-L, Zero-shot=true2022.05 | 27.3 | 49.6 | 58.3 | — |