We present a neural network-based upright frontal face detection syste
m. A retinally connected neural network examines small windows of an i
mage and decides whether each window contains a face. The system arbit
rates between multiple networks to improve performance over a single n
etwork. We present a straightforward procedure for aligning positive f
ace examples for training. To collect negative examples, we use a boot
strap algorithm, which adds false detections into the training set as
training progresses. This eliminates the difficult task of manually se
lecting nonface training examples, which must be chosen to span the en
tire space of nonface images. Simple heuristics, such as using the fac
t that faces rarely overlap in images, can further improve the accurac
y. Comparisons with several other state-of-the-art face detection syst
ems are presented, showing that our system has comparable performance
in terms of detection and false-positive rates.