# 2COOOL — Data

来源：https://www.kaggle.com/competitions/2coool/data

采集时间：2026-10-20T06:38:07Z

The 2COOOL Dataset combines data from three source datasets:

- The COOOL Dataset collected from the first COOOL Workshop - AlShami, A. K., Kalita, A., Rabinowitz, R., Lam, K., Bezbarua, R., Boult, T., & Kalita, J. (2024). Coool: Challenge of out-of-label a novel benchmark for autonomous driving. arXiv preprint arXiv:2412.05462.

- The DADA dataset. - Jianwu Fang, Dingxin Yan, Jiahuan Qiao, Jianru Xue, Hongkai Yu: DADA: Driver Attention Prediction in Driving Accident Scenarios.  IEEE Trans. Intell. Transp. Syst. 23(6): 4959-4971 (2022)

- The NEXAR Crash Prediction Dataset. - Moura, Daniel C., and Zvitia, Orly. "Nexar Collison Dataset." Hugging Face, 2025, https://huggingface.co/datasets/nexar-ai/nexar_collision_prediction.

If you use this dataset in your research please cite the data sources and follow the rules relating to attribution and redistribution found in our "Rules" page of the competition.

By downloading this dataset, you agree to the terms of this competition: [https://drive.google.com/drive/folders/1JYhSPv0zQQ_EBx2G4ljyF6CNXlxkRNVB?usp=sharing](https://drive.google.com/drive/folders/1JYhSPv0zQQ_EBx2G4ljyF6CNXlxkRNVB?usp=sharing)

2COOOL consist of ~670 short videos featuring rare hazards, crashes, and normal driving videos. The data labeling process was performed by four human annotation teams and later extended with additional data generated by VLMs and LLMs based off the human labels (with complete human review). We have also collected gaze data for each video to aid in hazard identification (found under the "heatmaps" folder). This provides a diverse view of potential incident report phrasings, diction, and content, enhancing the feedback teams will receive from the captioning tasks.

Each submission file should be formatted as follows, please ensure to **enclose all text fields with quotation marks to prevent evaluation problems:**

    video, Incident window start frame, Incident Detection, Crash Severity, Ego-car involved, Label, Number of Bicyclists/Scooters, Number of animals involved, Number of pedestrians involved, Number of vehicles involved (excluding ego-car), Caption Before Incident, Reason of Incident

    475, 110, 1, "4. Other cars collided with person/car/object but ego-car is ok", 0, vehicle drives into another vehicle, 0, 0, 0, 2, "Ego-car is driving through a city road with heavy traffic.", "A vehicle in front of ego-car tries turns left in front of oncoming traffic and collides with another vehicle."

The possible labels for each column are as follows:
```python
video: Numeric (0-671)
Incident window start frame: Numeric (1-total video frames)
Incident Detection: Numeric. 1=Accident, 0=Hazard, -1=No incident
Crash Severity: Textual [
"0. No Crash",
"1. Ego-car collided but did not stop", 
"2. Ego-car collided and could not continue moving", 
"3. Ego-car collided with at-least one person or cyclist", 
"4. Other cars collided with person/car/object but ego-car is ok", 
"5. Multiple vehicles collided with ego-car", 
"6. One or Multiple vehicles collided but ego-car is fine"
]
Ego-car involved: Numeric. 0=Not involved, 1=Involved
Incident Recognition: Textual [
"ego-car hits barrier"
"flying object hit the car"
"ego-car hit an animal"
"many cars/pedestrians/cyclists collided"
"car hits barrier"
"ego-car hits a pedestrian"
"animal on the road"
"car flipped over"
"ego-car hits a crossing cyclist"
"vehicle drives into another vehicle"
"ego-car loses control"
"scooter on the road"
"bicycle on road"
"pedestrian is crossing the street"
"pedestrian on the road"
"vehicle hits ego-car"
"ego-car hits a vehicle"
"vehicle overtakes"
"unknown"
]
Number of Bicyclists/Scooters: Numeric
Number of animals involved: Numeric
Number of pedestrians involved: Numeric
Number of vehicles involved (excluding ego-car): Numeric
Caption Before Incident: Textual (enclose in quotation marks)
Reason of Incident: Textual (enclose in quotation marks)
```