Mastering OpenCV 3 - Second Edition

Accounting for global geometric transformations

So far, we have assumed that the training images are centered at the facial feature and are normalized with respect to global scale and rotation. In practice, the face can appear at any scale and rotation within the image during tracking. Thus, a mechanism must be devised to account for this discrepancy between the training and testing conditions. One approach is to synthetically perturb the training images in scale and rotation within the ranges one expects to encounter during deployment. However, the simplistic form of the detector as a correlation patch model often lacks the capacity to generate useful response maps for that kind of data. On the other hand, the correlation patch model does ...