Hi there, I'm developing a face detector. Now I found a problem is that, the individual trained face detector works well on some evaluation dataset (AT&T face database & MIT non-face dataset) in terms of false postives (less than 1%) and false negatives (from 10%-1%). But when I apply some individual face detector on a set real-world images, the face detection rates vary from around 65% to 89%. In training I resize the AT&T face (a part of it, randomly selected 240 patterns) to 22 by 20 and use all 440 pixels for classifier training. Is there some way to improve the performance of the face detector, or to get a robust face dtector? What I could think of is that dimension reduction might be useful, or to increase the size of training set. Could anyone please give me some general suggestion? Many thanks. Bests, Q.H.Wang