### Scoring with mAP

Mean Average Precision (**mAP**) is used as the scoring function by evaluating the quality of temporal detection predictions. The model must identify the correct action label, along with the corresponding start and end timestamps within each video clip.

For each detection, the output should be in the following JSON format:

```json
{
  "video_id": ...,
  "t_start": ...,
  "t_end": ...,
  "label": ...
}
```

Predictions are evaluated against ground-truth annotations using **mean Average Precision (mAP)**, averaged over a set of temporal Intersection over Union (**IoU**) thresholds:

$$
\text{IoU} \in \{0.5, 0.75, 0.95\}
$$

For each IoU threshold, **Average Precision (AP)** is computed per class by ranking predictions according to confidence scores, computing precision-recall curves, and integrating the area under the curve. The **mean** is then taken across all classes and all specified IoU thresholds to yield the final mAP score.

This scoring method rewards both high localization precision (via IoU) and correct classification, and is well-suited for evaluating temporal detection tasks.

### Leaderboard Ranking

Final leaderboard ranking is based on the overall mAP score:

$$
\text{Final Score} = \frac{1}{|T|} \sum_{t \in T} \frac{1}{C} \sum_{c=1}^{C} AP_{c}^{(t)}
$$

Where:

- $T = \{0.5, 0.75, 0.95\}$ is the set of IoU thresholds  
- $C$ is the total number of classes  
- $AP_c^{(t)}$ is the Average Precision for class $c$ at IoU threshold $t$  

This metric encourages models to produce temporally precise and class-accurate predictions across a range of detection difficulties.