Image-to-Text Retrieval on COCO 2017 (val)
66Recall@1SmartCLIP
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| SmartCLIPModel Scale=L/142025.07 | 66 | 86.2 | 92.6 | |
| Long-CLIPModel Scale=L/142025.07 | 62.8 | 85.1 | 91.2 | |
| SmartCLIPModel Scale=B/162025.07 | 61.9 | 83.3 | 89.7 | |
| Long-CLIPModel Scale=B/162025.07 | 57.6 | 81.1 | 87.8 | |
| CLIPModel Scale=L/142025.07 | 56.1 | 79.5 | 86.8 | |
| CLIPModel Scale=B/162025.07 | 51.8 | 76.8 | 84.3 | |
| CLIP-RefineBackbone=ViT-B/32, Zero-shot=true2025.04 | 38.78 | 65.04 | 75.12 | |
| Direct Fine-tuningModel Scale=L/142025.07 | 37.9 | 63.1 | 72.2 | |
| HyCDBackbone=ViT-B/32, Zero-shot=true2025.04 | 37.88 | 62.54 | 72.84 | |
| Direct Fine-tuningModel Scale=B/162025.07 | 37.4 | 62.3 | 72.1 | |
| Self-KDBackbone=ViT-B/32, Zero-shot=true2025.04 | 35.36 | 59 | 69.72 | |
| HyCD + LalignBackbone=ViT-B/32, Zero-shot=true2025.04 | 35.14 | 61.74 | 72.06 | |
| Pre-trained (CLIP)Backbone=ViT-B/32, Zero-shot=true2025.04 | 33.26 | 59.1 | 68.78 | |
| m²-mixBackbone=ViT-B/32, Zero-shot=true2025.04 | 32.92 | 57.96 | 67.56 | |
| ContrastiveBackbone=ViT-B/32, Zero-shot=true2025.04 | 31.86 | 56.8 | 67.48 |