Face masks are now mandatory in the whole territory of Lombardia Area in the north of Italy and in many other areas worldwide. In an effort to contain the Covid-19 Coronavirus spread that has caused thousands of deaths, the local governments are insisting that millions of residents must wear protective face covering when they go out in public.
A funny consequence to covering their faces it’s the face masks trip up facial recognition functions, the technology necessary for many routine transactions in many country worldwide. Despite how much the face recognition is diffused a lot of experience such as certain mobile phones, entertainment apps and bank accounts won’t unlock anymore.
In January 2020, a Chinese company says it has developed the country’s first facial recognition technology that can identify people wearing a mask, as most are these days because of the #COVID19, and help in the fight against the disease.
The recognition rate can reach about 95%, which can ensure a quite good performance, the success rate for people without mask is estimated in 99.5%.
This Beijing-based firm used core technology developed over the past 10 years, a sample database of about a few million unmasked faces and a much smaller database of masked faces. It’s easy to argue how the new dataset has been inferred using a pretty easy to understand data augmentation technique: face with facial points, pick a random mask up, let’s project the mask according to the facial points = a (billion of) brand new masked faces.
In the same days someone teaches Iphone how to recognize his face wearing a mask. More: https://www.independent.co.uk/life-style/gadgets-and-tech/iphone-face-mask-id-recognition-unlock-coronavirus-covid-19-hack-a9459901.html
Most of the standard model in literature exploit information basically from eyes, mouth and face shape. And now? We starting figuring out how the face detection step should be modified / improved in order to handle with the missing information (occluded mouth and nose). So we developed the first LBP cascade, suitable for OpenCV, able to detect faces with masks. The cascade is totally free for academic, pro-bono or healthcare projects.
Masked faces LBP HAAR HOG OpenCV cascade
Full frontal (with partial profiles) masked face detection cascade, trained with:
- ~7,000 positive samples (randomly sampled)
- approx 0.9B of negative sub-regions containing outdoor and indoor samples (30%-70%)
- Training size w=40 h=50 (aspect ratio 0.8)
LBP: it’s free for academic, pro-bono or healthcare projects. Send us an email here to get the link for the cascade!
- Features set: 106080 features
- Training time: ~2 days
- TP: ~ 95.01% of positive training set
- FN: ~ 04.99% of positive training set
- FP: ~ 5.4e-006% of negative training set
HOG: (upon request)
HAAR: (upon request)