Loading…
Frame 0 / 23RGB Actual inference video
Loading frame…
Segmentation Ground-truth object colors
Loading frame…
Object correspondence
| Focus | GT | Ground-truth caption | GT visibility now | Pred. | Evaluation-matched predicted caption | Prediction now / display | Timing / positions |
|---|
Pairing is the saved evaluator assignment, not a caption-based rematch. Individual GT/prediction checkboxes persist while scrubbing. A matched caption can still be completely wrong.
Interactions and temporal errors
| Source | ID / type | Frame span | Inputs → outputs (captions) | Now |
|---|
Disable the active-frame filter to see out-of-range events. Ground ID 65535 in participant sets is normal; death=65535 is not.