Everything underneath

Frame corpus for AI training

JPEG stills from cameras on a working network, in taxi and rideshare cabins, at airport curbs, on casino and retail floors and at street-level displays. An index lists every still with its screen, capture time, market and venue, and each still our vision model described carries its one-line scene description and the model's whole response: the audience it read, what else it saw and its ad-fit verdict. Real lighting, weather, glare, motion blur and crowds, day and night.

Request a sample All datasets

The question it answers

Where do I get real in-cabin, curbside and venue imagery at scale?

4.2 million

stills

37,941

new stills a day in October

3.8 million

with a scene description

Measured over the corpus from May 5 to October 11, 2026.

The dataset

Grain: One image per capture.
Cadence: New stills every day. A delivery is one file per still and one index of them.
History: From May 5, 2026.
Format: JPEG images and a Parquet index.
Delivered to: Your S3 bucket, your GCS bucket, or a time-limited download link, with a manifest and a schema.
Method: How each measure is produced and how to add rows up: the measurement methodology.

What buyers do with it

In-cabin monitoring: occupancy, posture and seatbelt context in real cabins.
Perception for autonomous vehicles and robotaxis: curb, street and transit scenes.
Multimodal pretraining and evaluation, with image and text pairs out of the box.

The table

Read from the export registry each time this site is built.

`camera_frame_corpus`

One image per capture, from May 5, 2026

Grain: One row per call_id.
Delivery: One file per image, as JPEG images and a Parquet index.
Basis: Measured.
Definition: Version 3.

Column	Type	What it holds
`call_id`	string	The frame's identity: the cloud vision call it was captured for. The delivered JPEG is named by it.
`screen_id`	string	Trillboards screen id of the camera that captured the frame. Null: The call carried no screen id.
`called_at`	timestamp	When the frame was sent for analysis, UTC.
`captured_at`	timestamp	When the camera captured the frame, UTC. Null: Not recorded: the device shutter time is recorded from 2026-08-13; called_at is the nearest time.
`capture_mode`	string	What triggered the capture (e.g. periodic_scene, face_triggered). Null: Not recorded (before 2026-09-24, or captured off-device).
`capture_reason`	string	Why the device captured it (e.g. periodic_scene_baseline, face_detected). Null: Not recorded (as capture_mode).
`profile_id`	string	Sensing profile the call ran under (e.g. transit_hub_v1). Null: A legacy frame (the eval prefix records none).
`model`	string	Vision model that analysed the frame and wrote its caption. Null: A legacy frame.
`model_version`	string	That model's version. Null: Not recorded, or a legacy frame.
`object_key`	string	Path of the stored JPEG in our private media store (vision-frames-private/YYYY/MM/DD/<call_id>.jpg; vision-evals/ for the legacy frames).
`image_sha256`	string	SHA-256 (hex) of the stored JPEG, recorded when the frame was written. Null: Not recorded: the stored hash is written from 2026-08-20; the delivered file's own hash is in the index the rail writes.
`caption`	string	The model's description of the scene in the frame (its scene_description). Null: The call carried no scene description: a failed or truncated cloud call, a frame an edge model analysed on the device (model local-vlm), or a legacy frame.
`caption_response`	string	The model's whole response for the frame, as JSON text: its scene description with the audience classes, extra observations, social attention, ad-fit verdict and every other key it returned (2,853 characters on average over the 1,833 whole responses of 2026-10-10 12:00-13:00 UTC; the scene description alone averages 80). Null: The call's response is not whole JSON (a failed or truncated cloud call), or a legacy frame.
`country`	string	Country of the screen's registered location. Null: The screen has no registered country.
`state`	string	State or province of the screen's registered location. Null: The screen has no registered state.
`dma_code`	string	Nielsen DMA code of the screen's registered location. Null: Outside the US, or not geocoded.
`dma_name`	string	DMA name. Null: As dma_code.
`venue_category`	string	The screen's registered venue category (transit = a vehicle cabin). Null: Not classified.
`venue_type`	string	The screen's registered venue type. Null: Not classified.

Request a sample

Tell us the markets and the weeks you want. The sample is cut from production by the same export that makes the full delivery.

Request a sample

All datasets · Next: Place activity, hourly