We score solutions on the three things you predict for each clip: **accident time**, **impact location**, and **collision type**. Each part produces a score in **[0, 1]** (higher is better), and the final leaderboard score is their **harmonic mean** (so doing poorly on any one part hurts the overall score).

### Component scores

- **𝒯 (Temporal score)** — measures how close your predicted accident time is to the ground-truth time. It uses a Gaussian-style similarity: small time errors score near 1, larger errors drop smoothly toward 0.

- **𝒮 (Spatial score)** — measures how close your predicted impact point **(x, y)** is to the ground-truth location in the frame. It also uses a Gaussian-style similarity: small location errors score near 1, larger errors drop toward 0.

- **𝒞 (Classification score)** — **Top-1 accuracy** for collision type (1 if the predicted type matches, 0 otherwise; averaged across videos).

### Final score (harmonic mean)

$$
\text{ACCIDENT score}=\frac{3}{\frac{1}{\mathcal{T}}+\frac{1}{\mathcal{S}}+\frac{1}{\mathcal{C}}}
$$

## Submission file

Submit a CSV with a header and one row per test video:

```csv
path, accident_time, center_x, center_y, type
videos/iKpoAkiKqjw_00.mp4, 8.25, 0.633, 0.125, t-bone
videos/dQ0ao_Yu7II_00.mp4, 19.79, 0.197, 0.892, head-on
videos/nFpT2-RpAS8_2_00.mp4, 6.01, 0.529, 0.985, rear-end
videos/9AnFK85qRjM_00.mp4, 3.87, 0.847, 0.005, head-on
...
```