JackChew/egocentric-video-annotation-sample

Egocentric video annotation sample

Five short hands-on clips (cutting, folding, washing, pouring, stirring), each annotated by the LinkDigest video annotation API in one call. Every row is what the API returns with format: "annotation": timed frame descriptions, on-screen text, speech segments, key points and a coverage receipt that says what was not read.

• 5 clips · 148 seconds of video · 35 timed frame descriptions
• Read on 2026-11-07 through POST /api/v1/capture; a clip under a minute costs 2 credits (1 + 1 per started minute)
• Annotations: CC BY 4.0. Videos are not included. They are on Pexels under the Pexels License; each row links its source and author.

Clips

id	task	title	length	frames	video by
pexels-37239363	cutting	Slicing Red Onion in a Home Kitchen	33 s	8	Florian Delée
pexels-5116402	folding	A Person Folding a Bath Towel	47 s	11	Ian Panelo
pexels-8627110	washing	A Person Washing a Cooking Pan	17 s	4	Kampus Production
pexels-4794521	pouring	Pouring Hot Water on a Cup with Filter	32 s	8	cottonbro studio
pexels-4146100	stirring	Cooking Food in a Pan over the Stove	20 s	4	Natasha Filippovskaya

One row (frames and key points cut for display)

{
  "id": "pexels-37239363",
  "task": "cutting",
  "title": "Slicing Red Onion in a Home Kitchen",
  "video_url": "https://www.pexels.com/video/slicing-red-onion-in-a-home-kitchen-37239363/",
  "video_author": "Florian Delée",
  "video_author_url": "https://www.pexels.com/@florian-delee-209542985/",
  "video_licence": "Pexels License (video not included; see video_url)",
  "duration_seconds": 32.7,
  "read_on": "2026-11-07",
  "annotation": {
    "version": "1",
    "depth": "full",
    "speech": [],
    "frames": [
      {
        "t": 2,
        "description": "A hand holds a whole red onion on a yellow cutting board with vegetables in the background.",
        "ocr": ""
      },
      {
        "t": 6.1,
        "description": "A sliced piece of red onion rests on a yellow cutting board after being cut.",
        "ocr": ""
      }
    ],
    "on_screen": [],
    "chapters": [],
    "key_points": [
      {
        "text": "A red onion is being sliced on a wooden cutting board.",
        "quote": "red onion sliced",
        "t": null,
        "verified": false
      }
    ],
    "coverage": {
      "duration_seconds": 32.7,
      "images_read": null,
      "images_total": null,
      "largest_frame_gap_seconds": 4.1,
      "links_found": 0,
      "not_captured": [
        "no speech transcript: no speech was recognised in the audio",
        "comments not read"
      ],
      "seconds_transcribed": 0,
      "transcript_source": "none"
    }
  }
}

Fields

field	what it is
annotation.frames[]	t (seconds), description (what is in view, written by a vision model), ocr (text visible in the frame)
annotation.speech[]	start, end, text, approx (times estimated within the pieces the speech model hears). Empty here: all five clips are silent
annotation.on_screen[]	on-screen text with its start and end
annotation.key_points[]	a short summary point, the quote it rests on and whether that quote was found in the source (verified)
annotation.coverage	the receipt: seconds transcribed, the largest gap between frames, and not_captured in plain words

How it was made, and its limits

• The 640-pixel-wide file of each clip was read on one machine with ffmpeg: its sound (here silence, as the clips have none) and one frame every 4 seconds. Only those were sent; the video itself was not uploaded.
• Descriptions are model output, not human-checked. They describe single frames, so motion between frames is not described. There are no action or task labels, no hand or pose tracking and no depth.
• A clip with a voice gets timed speech segments as well; see the live sample on linkdigest.dev/annotation.

Annotate your own footage

The same read works on your own video: drop a file on the console's upload page (silent clips included), or pip install linkdigest-mcp and run linkdigest capture ./clip.mp4 --json. Price, limits and a free 10-minute pilot: linkdigest.dev/annotation.