Open-ended Video Question Answering on iVQA
39.6AccuracyFrozenBiLM
Evaluation Results
| Method | Links | |
|---|---|---|
| FrozenBiLMSupervision=100% (fully-supervised)2022.06 | 39.6 | |
| FrozenBiLMSupervision=10% (few-shot)2022.06 | 35.3 | |
| FrozenBiLMSupervision=1% (few-shot)2022.06 | 31.1 | |
| FrozenBiLMTraining Data=WebVid10M, Zero-shot evaluation=true2022.06 | 28.3 | |
| FrozenBiLMTraining Data=WebVid10M, Speech=true, Zero-shot=true2022.06 | 26.8 | |
| FrozenBiLMSupervision=0% (zero-shot)2022.06 | 26.8 | |
| FrozenBiLMTraining Data=WebVid10M, Speech=false, Zero-shot=true2022.06 | 26.2 | |
| Just AskTraining Data=HowToVQA69M + WebVidVQA3M, Speech=false, Zero-shot=true2022.06 | 13.3 | |
| Just AskTraining Data=HowToVQA69M + WebVidVQA3M, Zero-shot evaluation=true2022.06 | 13.3 | |
| CLIP ViT-L/14Training Data=400M image-texts, Speech=false, Zero-shot=true2022.06 | 9.2 | |
| CLIP ViT-L/14Training Data=400M image-texts, Zero-shot evaluation=true2022.06 | 9.2 | |
| RandomZero-shot=true2022.06 | 0.1 | |
| RandomTraining Data=—, Zero-shot evaluation=true2022.06 | 0.1 |