In “A GPU-Accelerated Face Annotation System for Smartphones,” Yi-Chu Wang, Sydney Pang and Kwang-Ting Cheng in the Dept. of Electrical and Computer Engineering, University of California, Santa Barbara, Ca., describe an OpenCV-based face annotation algorithm designed to run on an Android-powered Motorola Droid smartphone.
They used an OpenCV face detector algorithm to identify face regions in the given photo. Each face region is then aligned based on the face landmark detector and scaled down to a proper size.
For face recognition, they used a Gabor-based feature descriptor, which has been demonstrated as one of the most suitable local descriptors for face representation. This descriptor, derived by first convolving the face image with 40 Gabor wavelets of 5 different scales and 8 different orientations, exhibit desirable characteristics of spatial locality and orientation selectivity.
Their face recognition system achieved an accuracy of 90% on an uncontrolled family album dataset which consists of 18 people. For an implementation solely based on the CPU, it takes about 15 second to extract a face feature on a Droid phone.
When the embedded GPU on the Droid was used, the identification process was speeded up by 4 times bringing the overall response time to within their real-time requirement.