FaceDetection

Face Detection in Python using the Viola-Jones algorithm on the CBCL Face Database published by MIT's Center for Biological and Computational Learning. Learn how it works by reading my tuturial published in The Data Driven Investor on Medium.

Part one (The basic algorithm): https://medium.com/datadriveninvestor/understanding-and-implementing-the-viola-jones-image-classification-algorithm-85621f7fe20b

Part Two (The Attentional Cascade): https://medium.com/datadriveninvestor/understanding-and-implementing-viola-jones-part-two-97ae164ee60f

Code

• viola_jones.py
  • An implementation of the Viola-Jones algorithm
  • Viola, Paul, and Michael Jones. "Rapid object detection using a boosted cascade of simple features." Computer Vision and Pattern Recognition, 2001. CVPR 2001. Proceedings of the 2001 IEEE Computer Society Conference on. Vol. 1. IEEE, 2001.
• cascade.py
  • An implementation of the attentional cascade introduced by Paul Viola and Michael Jones
• face_detection.py
  • Methods to train and test a ViolaJones classifier on the training and test datasets
  • Methods to train and test a CascadeClassifier on the training and test datasets

Data

The data is described at http://cbcl.mit.edu/software-datasets/FaceData2.html, and I downloaded from www.ai.mit.edu/courses/6.899/lectures/faces.tar.gz and compiled into pickle files.

Each image is 19x19 and greyscale. There are Training set: 2,429 faces, 4,548 non-faces Test set: 472 faces, 23,573 non-faces

• training.pkl

  • An array of tuples. The first element of each tuple is a numpy array representing the image. The second element is its clasification (1 for face, 0 for non-face)
  • 2429 face images, 4548 non-face images

• test.pkl

  • An array of tuples. The first element of each tuple is a numpy array representing the image. The second element is its clasification (1 for face, 0 for non-face)
  • 472 faces, 23573 non-face images

Models

• 50.pkl
  • A 50 feature Viola Jones classifier
• 200.pkl
  • A 200 feature Viola Jones classifier
• cascade.pkl
  • An Attentional Cascade of classifiers looking at 1 feature, 5 features, 10 features, and 50 features.

Results

The hyperparameter T for the ViolaJones class represents how many weak classifiers it uses.

For T=10, the model achieved 85.5% accuracy on the training set and 78% accuracy on the test set.

For T=50, the model achieved 93% accuracy on both the training and test set.