Image-to-Text Relation Retrieval on Visual Genome Relations (top-50 predicate vocabulary)
3.48Top-1 AccuracyScenarioCLIP
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| ScenarioCLIPKnowledge Distillation=true, Evaluation=Zero-shot2025.11 | 3.48 | 13.17 | 20.18 | 4.3 | 13 | 19.23 | |
| ScenarioCLIP w/o KDKnowledge Distillation=false, Evaluation=Zero-shot2025.11 | 3.4 | 12.22 | 19.52 | 3.94 | 12.16 | 18.75 | |
| PyramidCLIPEvaluation=Zero-shot2025.11 | 3.36 | 12.5 | 19.58 | 4.25 | 13.04 | 19.07 | |
| CLIPBackbone=ViT-B/32, Evaluation=Zero-shot2025.11 | 0.8 | 3.55 | 5.08 | 1.61 | 4.58 | 8.63 |