When a global electronics maker came to DefinedCrowd with the goal of building more inclusive facial-recognition technology, the objective seemed simple: If given just one portrait containing a family of East-Asian descent, the model needed to be able to identify each individual person while also understanding that individual’s place within the family context. I.e. recognizing a little girl’s face as a little girl’s face as well as correctly identifying her as a “daughter.”
Training such a model was no easy task, and required accurately annotated images that adhered to highly specific criteria: Each portrait needed to contain children, have a minimum resolution of 640×640 pixels, and represent a wide variety of indoor lighting conditions.
Our globally-spread, actively managed Neevo workforce (45,000+ strong) was well up to this bespoke data collection. First, we sourced 20 unique pictures of 50 different families for a total of 1,000 images from our community, all verified by our team internally to ensure strict adherence to each and every one of the client’s specified parameters.
We used those collected images to set up a bounding-box task, where a new batch of Neevo contributors identified family members and provided their relationships, ages, and countries of origin (Father, 40, Japan for example). We had those labels verified and corrected by the crowd while RTA’s running and monitoring crowd behavior to ensure accuracy in the results.
As social media transitions from text to image and sensitive data is increasingly stored through digital channels, the need for facial recognition technology will only continue to grow. By partnering with DefinedCrowd, our client gained a leg up by obtaining an extremely customized dataset in just six weeks, half the timeframe for this kind of “high-touch” data collection offered by competitors.