Additional information for my poster session at PyCon 2017
As seen on mrnugget's excellent writeup on the topic, we can use the following script from his repo to generate additional positive images from a small sample pool of cropped positive images:
perl bin/createsamples.pl positives.txt negatives.txt samples 1500
"opencv_createsamples -bgcolor 0 -bgthresh 0 -maxxangle 1.1 -maxyangle 1.1 maxzangle 0.5 -maxidev 40 -w 80 -h 40"
There are a lot of options here that are better documented over at the repository, but the important bit is the 1500 which defines the quantity of images to generate. A good starting point is converting 40-60 positive images to 1500.
opencv_trainscascade comes bundled with OpenCV, and is the go-to tool for training these classifiers. Keep in mind that this can take a very long time. For instance, training in my example took about three days on a 2015 Macbook Pro.
opencv_traincascade -data classifier -vec samples.vec -bg negatives.txt -numStages 20 -minHitRate 0.999 -maxFalseAlarmRate 0.5 -numPos 1000 -numNeg 600 -w 80 -h 40 -mode ALL -precalcValBufSize 1024
-precalcIdxBufSize 1024
Again, we can see a lot of options here, and these are definitely worth playing around with on a per-case basis. The documentation can be found here.
OpenCV actually ships with a pretty cool sample script for real-time detection. It can be found at ~/opencv-3.1.0/samples/python/facedetect.py in your OpenCV installation. It takes a Haar classifier as a --cascade input, meaning it can be used with anything you've trained using this method (not just faces).