Visual Question Answering on VQAv2, OK-VQA, GQA (1k sampled test)
69.4VQAv2 AccInstructBLIP
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| InstructBLIPtraining split in model's training distribution=true2023.11 | 69.4 | 61.9 | 66.6 | 66 | |
| DRESSRegularization during RLAIF stage=true2023.11 | 62.8 | 67.8 | 58.9 | 63.2 | |
| LLAVA-HF2023.11 | 62.6 | 70.5 | 48.3 | 60.5 | |
| BLIP-22023.11 | 62.4 | 60.2 | 45.7 | 56.1 | |
| mPLUG2023.11 | 59.5 | 65.1 | 40.9 | 55.2 | |
| DRESSRegularization during RLAIF stage=false2023.11 | 59.1 | 58.9 | 50.1 | 56 | |
| LLaVA2023.11 | 58.5 | 63.1 | 48.8 | 56.8 | |
| miniGPT42023.11 | 53.7 | 58.1 | 40.2 | 50.7 |