The Aureus 3D Face Recognition Engine
Aureus FR uses the Aureus 3D pose correction algorithms under the hood without the need for a graphics card. It can handle large head pose, large expression changes and most lighting changes. It is re-entrant thread safe and can match more than a million templates per second on a single CPU.
It has a simple C style interface that can be called from any language capable of interfacing with the standard C bindings such as C, C++, C#, JAVA, Python, etc.
All tests performed using the color FERET dataset (http://www.nist.gov/itl/iad/ig/colorferet.cfm). Two frontal sets of images were used (set “fa” and “fb”) and eight non frontal sets were used (hl, hr, ql, qr, ra, rb, rc, rd) containing large head yaw (up to 70 degrees).
Simple to use: (1) Load (2) Initialize once (3) call “GetTemplate” and “MatchTemplates”
A rank test was performed in which the gallery was made from the two frontal sets and the probes from the remainder.
A similarity matrix was constructed in which every image was matched to every image and the results stored in a similarity matrix. The values along the diagonal of the matrix were the results of matching the images to themselves, they were therefore omitted from the ROC results. True Positives (TP), True Negatives (TN), False Positives (FP) and False Negatives (FN) were counted for a thousand different thresholds ranging from zero to unity. True Positive Rate (TPR) was calculated as TP/(TP+FN) and False Positive Rate was calculated as FP/(FP+TN).
Aureus 3D FR FERET Comps
Aureus 3D FR LFW Comps
Aureus 3D FR Spec Summary
Specifications for Aureus FR
Aureus FR is robust against:
- Pose: 60o yaw, 20o pitch and 30o roll automatically, (360o roll with manual initialization).
- Large expression changes
- Partial Face Occlusion
- Facial hair
- Lighting changes (excluding strong shadows)
- Size = 1200 bytes
- Generation time = 240ms using 3.0GHz CPU
- Match time = 0.8 micro seconds (1.25 million matches per second on single 3GHz CPU core).
- Template generation and matching are re-entrant thread safe.
- Face size (from chin to top of head) 90 pixels (equivalent to 30 pixel inter eye distance for front facing image).
- Grayscale dynamic range of 32 or more within face.
- Input image via function, can be grayscale, RGB, BRG, RGBA, BGRA with origin bottom left or top left.
- Windows 64 bit operating system.