Putting the smudginess of old VHS movies to use training facial recognition algorithms
Anyone who has learned to fly a plane knows that good pilots are not made in orderly classrooms. They are made in the disordered real world.
A team of Michigan State University researchers figured the same was true with facial recognition algorithms and unconstrained environments. The gap between unconstrained biometrics testing scenarios and semi-constrained training datasets of synthetic subjects hinders broader usefulness.
That gap is much larger when trying to train face biometrics from celebrities’ mugs found online.
The scientists have created a controllable face synthesis model that they say puts synthetic faces in the context of real-world systems. The model can make the images look like they were captured by CCTV cameras that are unlikely to be replaced for cost or other considerations.
It is a form of neural style transfer, in which one image style is transferred to an image. In this example, an ordinary photo of a dog is transformed into what the same dog might look like if painted by a specific artist.
Using the model, training images can be given sensor noise and motion blur, resolution can be minimized and the effect of turbulence can be created, according the researchers’ paper.
They state that one-to-one verification using an unconstrained training dataset is about a third lower than those using a semi-constrained dataset.
It is possible that a dataset holding tens of thousands of unconstrained subject images might address this problem, but the cost of manual labeling alone would be prohibitive.
The potential of synthetic data to solve challenges related to the ethical collection of data for facial recognition training has been prompted research, investment and excitement within the computer vision industry.