I've found 2 applications in IPP samples that use IPP face detection: Face_Detection and picnic.
1) What is the difference between them in performance? First processes images with more than 2**17 pixels in parts, the other does not differentiate larger images.
2) It uses flipped images around horizontal axis (that is bottom-up bitmap) and adjusts Haar cascade y coordinate accordingly. Is there specific performance reason to use bottom-up bitmaps instead of ordinary top-down?