The Prototype

I have created a searchable dataset inspired by the COCO explorable database to identify people from objects, specifically those that help with mobility or are used for medical reasons. There are approximately 17 photos in the database with the people and objects being separated from each other depending on what the focus is. In a fully imagined database, there would be significantly more images in the final project.

There are 4 different tags given to the photos in the database: Person, Wheelchair, Glasses and Walking Assist. These are all serchable in the database and are what an AI will categorise the images by.

I have linked the project below:

  • The Prototype

    • The Inspiration

    The COCO Dataset is the main inspiration for this prototype as it uses multiple images to teach an AI how to tell different objects apart. I noticed however that the database had no keywords for wheelchairs and crutches, something I found strange considering how important they are for many individuals. This means that while my prototype could be used individually, it could also be used to add to the COCO dataset to help with its reliability.

    I also noticed that COCO used shapes to identify objects. While this was fine for inanimate objects, when doing this with people, an AI may fail to notice specific attributes of a person, including the colour of their skin, possibly leading to a bias towards white people. Because of this, I used outlines to showcase the shapes of people so that smaller details could still be recorded by the AI in an attempt to prevent racist and coloured biases from occurring.

    I have linked my inspiration below:

  • COCO Dataset

    • Machine Learning

    The next step of this module would be to use machine learning to teach an AI how to distinguish between all these different images. I decided to create an extremely simplified version of this again in the teachable machine and uploaded my database to the system. Due to the limitations of Teachable Machine as a form of machine learning, this is not the method that would be used in the official practice. However, it is a start to visualising how an AI may begin to learn about these new types of object detection.

  • COCO Teachable Machine Model

    • What is the result?

    This prototype, imagined in full, would help machine learning to recognise medical objects such as wheelchairs as something to be tracked like humans. As I have learned during my time in the module "Digital Media and the Senses", AI can be biased against certain people and objects, meaning much important data will not be tracked as a result. This can reflect the world of digital media as media that uses motion detection and tracking technology such as the animation and gaming industries may struggle more to represent people who aren't able-bodied, causing morality issues within the industries. By attaching this prototype to a motion detector AI, applications such as Rokoko would be able to detect this vital equipment as part of the human being tracked, helping to solve the issue surrounding AI being biased and making the ways our bodies and senses interact with technology more accurate.

    © Untitled. Design: HTML5 UP.