Face detection using a cascade of boosted simple classifiers. Implemented from the Viola, Jones paper "Rapid Object Detection using a Boosted Cascade of Simple Features".
Vectorized NumPy code makes it pretty fast. In particular the function best_split_fast calculates the value of each split in a vectorized way, which seems to do around 10x speedup over a for loop.
Unfortunately, this still gives a lot of false positives. Longer training time and incorporating other types of simple features would probably help that. Also, there should be a better way to pick among overlapping detected windows.