Knowledge Retrieval on OK-VQA v1.1 (test)
89.32Recall@5RA-VQA-v2 (T5-large)
Evaluation Results
| Method | Links | |
|---|---|---|
| RA-VQA-v2 (T5-large)Base Model=T5-large, K=5, Knowledge Source=GoogleSearch, Retriever=FLMR2023.09 | 89.32 | |
| RA-VQA-v2 (BLIP 2)Base Model=BLIP 2 (T5-XL), K=5, Knowledge Source=GoogleSearch, Retriever=FLMR2023.09 | 89.32 | |
| RA-VQA-v2 (BLIP 2) w/o ROIBase Model=BLIP 2 (T5-XL), K=5, Knowledge Source=GoogleSearch, Retriever=FLMR2023.09 | 87.02 | |
| RA-VQA-v2 (BLIP 2) w/o ROI & VEBase Model=BLIP 2 (T5-XL), K=5, Knowledge Source=GoogleSearch, Retriever=FLMR2023.09 | 85.99 | |
| RA-VQA-v2 (BLIP 2) w/o ROI & Late-interactionBase Model=BLIP 2 (T5-XL), K=5, Knowledge Source=GoogleSearch, Retriever=DPR2023.09 | 83.43 | |
| RA-VQA-v2 (T5-large) w/o ROI & VE & Late-interactionBase Model=T5-large, K=5, Knowledge Source=GoogleSearch, Retriever=DPR2023.09 | 83.08 | |
| RA-VQA-v2 (BLIP 2) w/o ROI & VE & Late-interactionBase Model=BLIP 2 (T5-XL), K=5, Knowledge Source=GoogleSearch, Retriever=DPR2023.09 | 83.08 | |
| RA-VQA-v2 (BLIP 2) w/o Late-interactionBase Model=BLIP 2 (T5-XL), K=5, Knowledge Source=GoogleSearch, Retriever=DPR2023.09 | 82.9 | |
| RA-VQA (joint training)Base Model=T5-large, K=50, Knowledge Source=GoogleSearch2023.09 | 82.84 | |
| RA-VQABase Model=T5-large, K=5, Knowledge Source=GoogleSearch2023.09 | 81.25 |