Long-form Video Question Answering on EgoSchema EGO4D
67.2QA AccuracyHuman
Evaluation results
| Method | Links | |
|---|---|---|
| HumanObserved frames=1802023.12 | 67.2 | |
| TCRObserved frames=184*, higher proportion of patches dropped=true2023.12 | 35.1 | |
| TCRObserved frames=92x2, multiple forward passes=true2023.12 | 34.5 | |
| TCRObserved frames=922023.12 | 34.2 | |
| InternVideoObserved frames=8x11, multiple forward passes=true2023.12 | 32.1 | |
| BLIP2Observed frames=8x12, multiple forward passes=true2023.12 | 29.9 | |
| BLIP2Observed frames=82023.12 | 27.2 |