We do not directly provide the original video files. Researchers can download the original videos using the video IDs provided in the annotation files, rename them to yt_{Video ID}.mp4, and then use the annotation files along with our provided script scripts/generate_clips.py to extract all video clips. Specific usage instructions are provided within the script code.
The labels folder contains two video annotation files: labels_main.json includes annotations for 2,440 accident videos, and labels_safe.json includes annotations for 448 safe videos.
The annotation for each sample is formatted as follows:
{
"source_video": "yt_-8DwVUirUqE", // "yt_" + Original video ID
"filename": "yt_-8DwVUirUqE_1.mp4", // Clip's filename
"sport": {
"type1": "Standardized Sports Venues", // Environmental Setting
"type2": "Cycling" // Sports Category
},
"hazard": {
"level1": [
"Personal Factors" // Macro Inducing Factors
],
"level2": "Bicycle losing control after hitting a bump and crashing into a guardrail" // Direct Cause Description
},
"is_valid": true,
"absolute_start_frame": 0,
"absolute_end_frame": 212, // Annotation used for video cropping
"is_cropped": true,
"crop_box": {
"x": 1,
"y": 9,
"w": 702,
"h": 1254
},
"original_resolution": {
"w": 720,
"h": 1280
}, // Annotation used for video cropping
"fps": 30.0,
"timestamp1_frame_relative": 149, // Earliest Cue Moment
"timestamp2_frame_relative": 167 // Most Obvious Moment
}