Basic Sliding Window Detector
The basic sliding window detector trains an SVM, either linear or nonlinear, on positive
and negative examples of faces.
At test time, the detector gathers patches of each image and tests each patch to see whether
it is a face. Patches are centered at points on a grid with STEP_SIZE between samples, across
grids of multiple scales spaced SCALE_STEP in scale space.
The default behavior is to use random negative crops in the training process. This yields
detectors with a high variance of performance. For example, using HOG features at the
default STEP_SIZE and SCALE_STEP, multiple runs with the same parameters yielded final
AP scores between .2 and .4 . This is undesirable in a model.
Also, it should be mentioned that in order to achieve Total Recall, it was necessary to
push the bias of the model up by 1000 or so.