You need to agree to the conditions to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

Social Gaze Annotations

Person-specific, frame-level gaze target annotations for multi-person social videos.

  • Videos: 133 (83 One Night Ultimate Werewolf game sessions, 50 YouTube social-interaction videos)
  • Frames: 216,904 at 5 fps (~12 hours)
  • People: 1,000 person tracks
  • Annotators: 3 independent annotators per video

Annotation

Each person is tracked with an ID-consistent face box. For every frame, each annotator either marks where that person is looking or gives a reason why the target cannot be marked.

Field Meaning
id Person ID, consistent within a video
face_bbox Face box [x1, y1, x2, y2] in pixels (640×360 frames)
gaze_point Gaze target point [x, y] in pixels, when marked
target_person ID of the person being looked at, or None for non-person targets
problem None if marked, face if the person's face/eyes cannot be read, target if the target is not visible

Layout

round1_werewolf/  round2_youtube/  round3_werewolf/
  op1/ op2/ op3/   # one annotator each: <video>.mp4.json
  merged/          # three annotators combined (comma-separated values)

Each JSON file is a list of frames:

{"frame": 1, "segments_info": [{"id": "0.1", "face_bbox": [...],
  "attributes": {"gaze_point": [x, y], "target_person": "0.2", "problem": "None"}}]}

Notes

  • frame is 1-based on the 5 fps video.
  • In round 3, the problem key is omitted for marked points.
  • Source videos are not included.
Downloads last month
25