These pictures are from a well-known face database, "Labeled Faces in the Wild". This database provides benchmarking tests based on pair matching to compare performances between algorithms. And the pictures are collected from news, so each face image is taken under unconstrained conditions. Unconstrained conditions involve large variations caused by pose, lighting condition, facial expression, occlusion, age, blur, image quality and so on. As you can see, images from the same person may look quite different and that is the main reason of the difficulty.
|
This table represents results of commercial and current best performing systems on LFW benchmarks (face verification perforamance). Here, ‘face.com’ is one of the most successful companies about face recognition. They have supplied web-based API for face recognition through internet until recently, and have been acquired by Facebook on June 2012. As you can see, the world-best performance is a little more than 90% in case of the unconstrained face verification. The performances are closing to human accuracy where humans achieve around 97%. Then, how about in case of unconstrained face identification?
| |
This graph is answering the question. In the paper of ‘face.com’, they compared their method with others under unconstrained face identification using LFW dataset. The recognition rate is started from 80% at 5 classes and decreases by up to 40% at 100 classes. This result clearly shows the current limitation of face recognition.
|
This flow diagram is about general steps of unconstrained face identification. From detected face region, face is aligned using general alignment method or piece wise affine warping with facial feature points from facial feature extractor. Before face description step, cropping face into components can maximize benefits from localization. And also, some compensation like illumination and pose compensation may be needed to reduce variations caused by unconstrained conditions. And then face is represented into vector or matrix using face descriptor. At that step, codebook or dimensional reduction scheme may be used as necessary. Finally, the represented face is passed through trained classifier to get final result. The following figure shows the demonstration of the flows.
|