TR2007-074

A New Weak Learning Algorithm for Real Hyperplane Features Applied to Face Detection


    •  Raphael Pelossof, Michael Jones, "A New Weak Learning Algorithm for Real Hyperplane Features Applied to Face Detection", Tech. Rep. TR2007-074, Mitsubishi Electric Research Laboratories, Cambridge, MA, June 2008.
      BibTeX TR2007-074 PDF
      • @techreport{MERL_TR2007-074,
      • author = {Raphael Pelossof, Michael Jones},
      • title = {A New Weak Learning Algorithm for Real Hyperplane Features Applied to Face Detection},
      • institution = {MERL - Mitsubishi Electric Research Laboratories},
      • address = {Cambridge, MA 02139},
      • number = {TR2007-074},
      • month = jun,
      • year = 2008,
      • url = {https://www.merl.com/publications/TR2007-074/}
      • }
  • MERL Contact:
  • Research Areas:

    Artificial Intelligence, Computer Vision, Machine Learning

Abstract:

This paper explores the use of thresholded hyperplanes as the building blocks of a classifier for face detection. We are motivated by the work of Viola and Jones [10] who used Haar-like wavelet features as their weak classifiers in the AdaBoost learning algorithm. These weak classifiers were chosen for their speed. We explore how much may be gained by using more powerful but less computationally efficient weak classifiers. The generalized haar wavelets used in Viola and Jones can be viewed as a constrained subset of linear hyperplanes. Can a more powerful detector be constructed if we use unconstrained linear hyperplanes in place of the generalized Haar wavelets. In addition to being of theoretical interest, this question has practical importance for hardware implementations of a face detector in which dot products may be very fast to compute. The difficulty with using thresholded hyperplanes as weak classifiers is that the brute force search over all possible hyperplanes which was used in Viola-Jones is no longer practical. We propose a new gradient descent based algorithm which finds separating hyperplanes by directly minimizing the AdaBoost Z score. We also provide a baseline comparison to other search algorithms for unconstrained hyperplanes.