Face Detection Algorithm and Feature Performance on FRGC 2.0 Imagery

The performance of three well known face detection algorithms and four alternative types of features are characterized using face data from the Face Recognition Grand Challenge. The three algorithms are a semi-naive Bayesian classifier, a neural network called a SNoW, and a cascade classifier using Haar wavelets. For the first two algorithms, ROC analysis is used to assess the relative value of wavelet features compared to simpler pixel features. No universally best feature is observed, and for imagery acquired under uncontrolled lighting, pixels perform slightly better than wavelets. The cascade classifier is found to be impossible to train in the same fashion as the other algorithms, but it is also found to perform very well using a training configuration supplied along with the algorithm as part of the OpenCV library.