Datasets:

Yogesh914
/

RiskCueBench

license: apache-2.0
configs:
  - config_name: default
    data_files:
      - split: carcrash
        path: data/carcrash-*
      - split: protest
        path: data/protest-*
dataset_info:
  features:
    - name: video_id
      dtype: string
    - name: riskVisualIndicator
      dtype: string
    - name: riskSignalDescription
      dtype: string
    - name: riskSignalStart
      dtype: string
    - name: riskSignalEnd
      dtype: string
    - name: accidentStartFrame
      dtype: string
    - name: accidentEndFrame
      dtype: string
    - name: riskLabel
      dtype: string
  splits:
    - name: carcrash
      num_bytes: 79399
      num_examples: 502
    - name: protest
      num_bytes: 103623
      num_examples: 484
  download_size: 76908
  dataset_size: 183022
task_categories:
  - video-classification
  - video-text-to-text
language:
  - en
size_categories:
  - n<1K

🚨 RiskCueBench

A benchmark dataset for evaluating risk reasoning capabilities in video understanding models

📖 Overview

RiskCueBench provides fine-grained annotations of risk signals—visual cues that precede potentially dangerous events. This dataset contains annotated video clips from two domains: 🚗 traffic accidents and 📢 protest events, designed to test temporal risk anticipation and visual reasoning.

✨ Key Features

	Feature	Description
🎯	Risk Signal Annotations	Temporal boundaries marking when risk indicators appear
📝	Rich Descriptions	Detailed narratives of visual cues and event progressions
🏷️	Binary Labels	Clear yes/no labels for whether risk materializes
🌐	Cross-domain	Two distinct domains for generalization testing

🎓 Evaluation Tasks

The dataset enables evaluation of models on:

🔍 Risk Signal Detection — Identifying visual cues that indicate potential danger
⏱️ Temporal Reasoning — Understanding the progression from risk signals to outcomes
🔄 Cross-domain Generalization — Testing on both traffic and social scenarios

📊 Dataset Statistics

Split	Domain	Samples	Description
`carcrash`	🚗 Traffic	502	Dashcam footage of driving scenarios
`protest`	📢 Social Events	484	Protest and crowd footage

📥 Downloading the Videos

🚗 Car Crash Videos

The car crash videos are sourced from the Car Crash Dataset (CCD) published in ACM MM 2020.

To download:

🔗 Visit the official repository: CarCrashDataset
📦 Download the dataset from Google Drive (link provided in the repository)
🔢 The video_id column corresponds to the video filenames in the Crash-1500 and Normal folders

📚 Citation for CCD

@InProceedings{BaoMM2020,
    author = {Bao, Wentao and Yu, Qi and Kong, Yu},
    title  = {Uncertainty-based Traffic Accident Anticipation with Spatio-Temporal Relational Learning},
    booktitle = {ACM Multimedia Conference},
    year   = {2020}
}

📢 Protest Videos

The protest videos are sourced from YouTube. The video_id column contains the YouTube video ID.

URL Format:

https://www.youtube.com/watch?v={video_id}

Example: For video_id = "5gM1gnMkUKU" → https://www.youtube.com/watch?v=5gM1gnMkUKU

💡 Tip: Use tools like yt-dlp or pytube to download videos programmatically.

📋 Column Descriptions

Column	Type	Description
`video_id`	`str`	🎬 Unique identifier for the video. For car crash: filename ID from CCD. For protest: YouTube video ID.
`riskVisualIndicator`	`str`	👁️ Concise description of visual cues that signal potential risk
`riskSignalDescription`	`str`	📝 Detailed narrative of what happens during the risk signal period
`riskSignalStart`	`str`	⏱️ Frame marking the beginning of the risk signal
`riskSignalEnd`	`str`	⏱️ Frame marking the end of the risk signal
`accidentStartFrame`	`str`	💥 Frame when incident begins (`"/"` = no incident)
`accidentEndFrame`	`str`	🏁 Frame when incident ends
`riskLabel`	`str`	🏷️ `"yes"` = risk materializes, `"no"` = remains safe

💻 Usage Example

from datasets import load_dataset

# 📦 Load the full dataset
dataset = load_dataset("Yogesh914/RiskCueBench")

# 🔀 Access specific splits
carcrash_data = dataset["carcrash"]
protest_data = dataset["protest"]

# 🎯 Filter for risky scenarios
risky_carcrash = carcrash_data.filter(lambda x: x["riskLabel"] == "yes")
risky_protest = protest_data.filter(lambda x: x["riskLabel"] == "yes")

# 📄 Access a sample
sample = carcrash_data[0]
print(f"Video ID: {sample['video_id']}")
print(f"Risk Signal: {sample['riskSignalDescription']}")
print(f"Risk Label: {sample['riskLabel']}")

📝 Data Samples

🚗 Car Crash Example (Risk = ✅ Yes)

video_id: 1
riskVisualIndicator: "Black car, intersection"
riskSignalDescription: "First, a black-colored car enters the intersection against 
                        the traffic, then it continues into the driver's path."
riskSignalStart: 23        # frame number
riskSignalEnd: 31          # frame number
accidentStartFrame: 32     # frame number
accidentEndFrame: 50
riskLabel: "yes"

📢 Protest Example (Risk = ❌ No)

video_id: "5gM1gnMkUKU"
riskVisualIndicator: "Police body language, repeated hand gesture"
riskSignalDescription: "First a group of geared police show up, then one police 
                        use subtle body language to express message seem to be 
                        'come here', and did it twice."
riskSignalStart: "00:00"   # MM:SS format
riskSignalEnd: "00:13"     # MM:SS format
accidentStartFrame: "/"    # no incident
accidentEndFrame: "/"
riskLabel: "no"

📢 Protest Example (Risk = ✅ Yes)

video_id: "m6CwOP4rUAo"
riskVisualIndicator: "Police tank appears, gun raised, aiming gesture"
riskSignalDescription: "First a few police on police tank show up, then the 
                        police tank raise gun up and target at the protestors"
riskSignalStart: "00:17"   # MM:SS format
riskSignalEnd: "00:18"     # MM:SS format
accidentStartFrame: "00:19"
accidentEndFrame: "00:25"
riskLabel: "yes"

📜 License

Dataset	License
🚗 Car Crash (CCD)	MIT License — Original Repo
📢 Protest Videos	Subject to YouTube Terms of Service

📚 Citation

If you use this dataset, please cite:

@dataset{riskcuebench2025,
    title = {RiskCueBench: A Benchmark for Video Risk Reasoning},
    author = {[Authors]},
    year = {2025},
    publisher = {},
    url = {https://huggingface.co/datasets/Yogesh914/RiskCueBench}
}

Additionally, please cite the Car Crash Dataset if you use the carcrash split:

@InProceedings{BaoMM2020,
    author = {Bao, Wentao and Yu, Qi and Kong, Yu},
    title  = {Uncertainty-based Traffic Accident Anticipation with Spatio-Temporal Relational Learning},
    booktitle = {ACM Multimedia Conference},
    year   = {2020}
}

Made with ❤️ for the research community