Talk to us

Vision data

In development

Image data from people who were actually there

AI teams need current images of streets, shops and buildings. We are building tasks where players capture those images on site, in 70 countries.

This product is in development. This page describes what we are building and what already exists underneath it.

Place
Grote Houtstraat, Haarlem
Device
Attested
Privacy
Faces removed
Rights
Licence on record
Illustration on a public domain street photo: what a capture should look like on delivery. Photo: Fons Heijnsbroek, CC0

The gap

What is missing today

01

It goes stale

Street-level imagery is refreshed every few years at best. A shopfront, a road layout or a building site can change in a week.

02

It follows the car

Mapping vehicles cover roads. Footpaths, markets, stations and much of the world's cities are thin or missing.

03

It cannot be vouched for

Scraped images come with no proof of where or when they were taken, and no clear consent from whoever took them.

How it will work

How it will work

01

You define the task

Places, angles, conditions

A list of storefronts, a set of crossings at night, one street in the rain. You say what the model needs to see and how much of it.

02

It becomes a quest

On the map, for players nearby

The task appears in the app for the people who are already close, with a reward for completing it.

03

People capture it on site

Verified place and time

A capture only counts from an attested device inside the radius, so every frame carries a place and a time that were checked.

04

You receive a dataset

Reviewed and privacy filtered

Checked for quality, filtered for privacy, and delivered with its location and time metadata.

From capture to training

What happens to a capture

A photo is not training data yet. Five steps sit between the capture and your model.

  1. CaptureOn site, from an attested device.
  2. VerificationThe server checks place and time.
  3. Privacy filterFaces and number plates are removed.
  4. LabelsObjects get a label and a box.
  5. DeliveryAs a dataset, with metadata and a licence.

Labelled crops, as they would sit in a dataset

  • sign
  • sign
  • facade
  • banner
Illustration: crops from the same public domain street photo.

Uses

What this kind of data is used for

A model only knows the physical world from pictures of it. That holds for an assistant that answers questions about a photo as much as for a robot crossing a square. Here is one street, and eight things a model can take from it.

Illustration: the layers are placed by hand on a public domain street photograph. No model produced them.

Already in place

What already exists

The capture product is new. The network, the verification and the quest engine are already running.

A live networkPlayers in 70 countries, 31 languages
Verified presenceAttested devices, mocked locations refused
A quest engineMulti-step quests, reviewed and staged
Reward railsPayouts with daily caps and review
Vision modelsAlready reading street photos for our 3D map

Principles

Our ground rules

Consent and payment

Contributors know what they capture and why, agree to it, and are paid for it.

Privacy before delivery

Faces and number plates are to be removed before any dataset leaves us.

Clean rights

Every image is to come with a clear licence from the person who took it, traceable to a verified capture.

Design partners

What images does your model need?

We are shaping the first capture tasks with a small number of AI teams. If you have a data gap in the physical world, we want to hear what it is.