Streaming metrics and episode filters use the revised stream history. The live lane, when available, shows predictions emitted at each step before revision.
Failure modes — per annotation tag
episodes is how often a run carrying this tag comes out perfect end to end;
frames is how often the decode labels an individual tagged frame the way the data does.
A tag can look fine by frame and destroy the episode. An episode carrying two tags is counted
under both, so the rows do not sum to the split.
Episodes
j/k walk · space replay
· clicking a lane or a missed stage seeks the video
Pick an episode.
Stages the model missed
Segments below the IoU bar, worst first. step is how far the labels moved
from the previous stage — +1 is ordinary, anything else is the annotation skipping
or reversing. off by is how far the model sat from the label over those
frames. Click a row to jump the video there.
Progress
label
window
stream