Revolutionizing Inclusivity: A New Era for Facial Recognition Technology
Artificial intelligence is constantly evolving, and facial recognition technology is no exception. However, current systems tend to exhibit biases, performing well with some populations but struggling with others. Addressing these shortcomings is crucial as the demand for more accurate and inclusive facial recognition technology grows. In this success story, we will discuss how a global electronics maker is taking bold steps to develop a more inclusive facial recognition model, ensuring that the technology works equally well for all populations.
Our client, a holding company that operates through its subsidiaries, is a leader in providing electronics and connectivity solutions for worldwide customers.
Although facial recognition technology is improving exponentially, it is only somewhat successful in recognizing people from different demographics. When identifying the faces of white males, the algorithms perform admirably. However, when attempting to recognize male or female African American and Asian faces, its accuracy drops significantly. It seems real-world bias has affected the functioning of AI machines.
However, AI software is only as “clever” as the data it trains on. The more inclusive the data, the more inclusive the technology. This salient point was recognized by a global electronics maker who wanted its software to decipher a single photograph of an East Asian family correctly. By developing more inclusive facial recognition technology, they aimed to tackle the existing biases in the industry.
This new model needed to identify each individual in the photographs provided while also understanding an individual’s place within the family context. In other words, the software required recognizing a little girl’s face as a little girl and correctly identifying her as a “daughter”. Training such a model required accurately annotated images that adhered to precise criteria: each portrait needed to contain children, have a minimum resolution of 640×640 pixels and represent a wide variety of indoor lighting conditions. This would ensure that inclusive facial recognition technology could perform consistently in different settings.
Step 1 – Image collection
With a crowd of over 500,000 people spread across 195 countries, we were able to source 20 unique images of 50 different families, resulting in 1000 images. To ensure strict adherence to the client’s specified parameters, each image was carefully verified by an internal team. This diverse dataset was critical for developing inclusive facial recognition technology.
Step 2 – Image tagging
The next step was to label the collected images. Our crowd annotated each image through bounding-boxes (a classification system in which annotators draw a box over the object of interest, based on the client’s requirements). In each photograph, annotators identified family members and provided information on their relationships to each other, their ages, and their countries of origin (father, 40, Japan, for example).
Our team verified and corrected each label while running real-time audits (RTAs) and monitoring crowd behavior to ensure accurate results.
As social media transitions from text to image and sensitive data is increasingly stored through digital channels, the need for inclusive facial recognition technology will only continue to grow. To keep our clients at the forefront of this developing technology, we provided them an incredibly customized dataset in just six weeks – half the timeframe competitors offer. This rapid progress showcases the potential of inclusive facial recognition technology and its impact on the industry.
Embrace the future of Computer Vision and stay ahead of the curve. Check our ready-to-use solutions on our Marketplace or contact us today to learn more about our cutting-edge solutions and how they can benefit your organization.