You always scale the original image down, it is pretty standard

The point being made here isn't how to get face recognition speed workable on a pi though it's that this is no faster than the common implementations despite the catchy title.

Yes but not to 128x96. Usually you want at least 640x480.

Wrong. Most facial recognition software reduces largest image dimension to 128. This allows caching to become effective, massively speeds up the process (hence the 15k fps).

A facial recognition pipeline contains a few steps, for feature extraction people DO resize image to something like 128x128, but for face detection (which is what this repo does) you DON'T.

Also, the reason why we scale down the input image is not for cache effectiveness, it's simply for reducing computations needed. Maybe except for naive matrix multiplication and convolution implementations, which is what this repo does. But there is no point to discuss performance if you are using a naive implementation which by design ignores cache/instruction latency/anything Computer Architecture related. Please, at least take QNNPACK as baseline.

Seems like something one might use in a pipeline following a Haar cascade to detect faces in a larger image in some sort of surveillance camera scenario.

I don't think that would be helpful, haar cascades have too many false-negatives (in part because it relies too heavily on specular reflections from your face's T-zone). Probably worth just using the CNN, you can make something else do region proposals if you really want. Probably another net :)