I. Background

This is my implementation of a face detector based on the paper
published by Viola and Jones in 2001 entitled Rapid Object Detection
using a Boosted Cascade of Simple Features.

Original paper:
https://www.cs.cmu.edu/~efros/courses/LBMV07/Papers/viola-cvpr-01.pdf
Wikipedia:
https://en.wikipedia.org/wiki/Viola%E2%80%93Jones_object_detection_framework

II. Prerequisites:

1. python2.7
2. numpy
3. python imaging library
4. matplotlib (optional)

This project comes with an optional C++ module that significantly
accelerates training and classification. Compiling this module has
additional prerequisites:

1. GCC (if you want to compile with visual studios on windows,
   you'll need to fiddle with setup.py)
2. numpy headers (I believe this come with numpy by default?)
3. python headers (apt-get install python-dev on debian/ubuntu)

Compile the C++ module by running the script build.sh.

Troubleshooting tip: If you have problem with errors during
compilation and have a different version of GCC (I'm using 4.6.2)
trying removing the -Wall and -Werror flags from setup.py.

III. Running:

driver.py is used to execute all functionality.

python driver.py --help # for help

Examples:

# Run on tiny dataset.
python driver.py

# Same, but use c++ backend.
python driver.py -c

# Read in 3000 faces, 3000 non-faces. select 5000 for training and 1000 for validation.
# Do boosting with 50 iterations. This will take hours.
python driver.py -c -f 3000 -o 3000 -s 5000 -v 1000 -i 50 

IV. Saved Classifiers

Training on large datasets can take quite a while, so I've prepared
some classifiers pretrained on large datasets (5000 images).

large_classifier.json
new_classifier.json

you can load these quickly and run them against a validation set with:

python driver.py -c -f 3000 -o 3000 -v 6000 --load-classifier new_classifier.json

This will be much faster than training a large dataset
yourself. Loading all the faces into memory will still take a couple
of minutes though.