Results on the sample videos

What the model finds, and what it misses

Scores are computed with the organisers' evaluate.py on our own labels of the four sample videos (dev_labels.json), using the frozen submission output predictions_samples.json. The organisers' hidden test set will score differently.

Annotated sample videos

Rendered by our own tooling from the tracks: boxes with track ids, scene geometry (stop line red, solid lines yellow, crossings white), the signal state read from the lamp, active events and the Part B risk bar. The timeline shows predictions against our labels.

Scores per class

Mean F1 over tIoU 0.3 / 0.5 / 0.7

Score A is the mean of these bars across every class in labels ∪ predictions. A labelled class we never predict (congestion) counts as 0.

Examples of each detected class

For each class, the prediction with the highest temporal IoU against our labels. Click to play it.

Failure cases

For each class, the longest false positive and the longest missed label at tIoU 0.5, with what went wrong.

Operator view

What a traffic centre would see across the four clips: how often each violation happens and for how long.

Events per hour of footage, by class

Events per clip

Ablation