Bootstrapping a global AI training dataset for the visually impaired
A Microsoft-funded research project is preparing to expand an AI program intended to help blind people and those with low vision to find belongings in familiar, cluttered surroundings. It is object recognition, but the objects are specific to a person.
City, University of London already has been gathering relevant data for personalized object recognition for its Object Recognition for Blind Image Training, or Orbit, research project.
The first phase has occurred in the UK, according to a Microsoft post, and involved 48 visually impaired subjects who uploaded more than 3,000 videos. Data collection is expected to go global this month.
Taking part in the project is not an effortless endeavor for anyone involved. Creating a representative dataset could seem tedious for some with no promise of a working product in the near future.
Volunteers use an iPhone to shoot short videos of at least five objects in a process researchers say should take no more than an hour before they are automatically uploaded.
Each object is recorded in video several times, at least one of which must be in an overall setting where the object usually is — coat by a door, for example. Another video must show the object alone on a cleared surface.
Participants also have to create reference videos using their hands as the reference, and moving the phone toward and away from the object.
Research staff will judge each clip, looking for identifying information and inappropriate content.