The Aureus 3D SDK has a rich feature set that will empower your facial recognition infrastructure. Conceptually Aureus 3D is very simple – give it an image, make a 3D model of the face seen in the image and output a new, optimized image. When you add Aureus 3D into your tool chain you’ll see immediate results. From a reduction in your Failure to Enroll Rate and False Rejection Rate to a huge jump in real-time monitoring effectiveness the benefits to having Aureus 3D in-line will become as clear as the nose on your face!
The previous version of this technology set the standard for single image 2D to 3D reconstructions. In order to leapfrog our own technology we decided the next problem address would be to find a way in which facial recognition could be made useful in sureveillence situations. With that in mind, Aureus 3D was developed to be a system that could locate, track and convert to 3D any person in an image sequence.
The most important part of such a system is its ability to accurately generalize to different identities in different environments. There are four major sources of variation that a working system must accommodate: (a) identity, (b) environment, (c) head pose and (d) expression.
Aureus 3D utilizes two highly advanced 2D face location and tracking algorithms. Of these algorithms, one is used to locate faces in incoming frames; the second (a neural network) is used as a failsafe mechanism to ensure that a currently tracked face is really a face. For this generation of software, Aureus 3D’s facial feature location and tracking algorithms have been improved in four ways; (a) developing scale invariant models, (b) utilizing the I3 de-correlated color channel transformation, (c) developing a shape only model capable of rapidly re-locating facial features from a greater distance (which is essential in sequences containing camera shake) and (d) constructing a posed set of models each of which can deal with their own pose range.
Aureus 3D’s head reconstruction covers a wider 3D pose range and allows an increase in 3D texture resolution and noise reduction. The multi-pose extruder has been extended from +/- 35 degrees yaw to +/-70 degrees yaw. Three different texture resolutions can now be selected from (low, medium and high). Additionally, the use of posed texture map masks has been implemented to improve the resulting 3D texture.
When Aureus 3D was to finally compiled, we tested the reconstructed 3D heads with the best commercially available 2D facial recognition. Experiments were designed and performed to provide a comparison between a commercially available facial recognition system and a simple parametric recognition algorithm. It was determined that the parametric identification method produced similar identification but over a larger pose range. An additional experiment was performed to determine the possible increase in identification that can be obtained by correcting for 3D head pose via the reconstructed 3D head meshes and applying the commercial recognition system. An average rank zero identification improvement of 28% was obtained over a pose range of +/-70 degrees yaw and +/-25 degrees pitch.