Unconstrained Facial Recognition

Facial recognition technology has been sitting in the attic for a long time, since it worked only for idealistic conditions. These idealistic conditions correspond to a mugshot like frontal view of the face in a neutral expression with well lit ambient, diffused light on the faces. These constraints more or less prevented it from being used in realistic scenarios where none of the above constraints hardly ever holds true.

In real life, people turn their heads around, make different expressions, and the illumination can range from dark shadows to them being in the limelight (literally!). Moreover, their appearance changes with age, facial hair and/or facial decorations such as sunglasses or makeup. Welcome to the unconstrained world of facial recognition, where life is much more harder but very interesting!

Broadly speaking, there are four confounding factors for facial recognition in the unconstrained setting — Age, Pose, Illumination and Decorations (A-PIED). Therefore, in order to be applicable to everyday scenarios, facial recognition must be A-PIED invariant.

Traditional facial recognition techniques cover a very small pie of the A-PIED invariance and hence fail miserably. In this case, the realistic images are from a different A-PIED slice than the enrolled images. With the advent of deep learning, the A-PIED slice has become considerably bigger.

At VCognition, we have developed deep-learning based facial recognition technology that can represent identities of people from a very small collection of images (even a single image works!) of a person. Our technology harvests the power of deep neural networks in the backend to expand the A-PIED slice. We have been recently ranked at #2 in NIST IJB-A benchmarks where the images are representative of realistic in-the-wild unconstrained conditions.

Stay tuned for our next post on accuracy and speed tradeoffs for large scale facial recognition technology.