Face and Automobile Detection using Statistical Modeling

Henry Schneiderman and Takeo Kanade
Robotics Institute
Carnegie Mellon University
Pittsburgh, Pennsylvania


To develop accurate object recognition algorithms.

We wish to detect upright human faces that vary in orientation from profile to frontal view.

For automobiles we are limiting ourselves to 2 door and 4 door passenger cars and a restricted range of view points.

For both problems we place no restrictions on background scenery, lighting, or other natural environmental conditions.


  • Distinguish the object from anything else we may encounter in the world.

  • Accomodate variation in the appearance of an object.

    Variation of the object itself:


    Variation due to lighting:

    Variation due to pose (the geometric relationship of the camera to the object):



    Our approach has 3 components:

    1.   Statistical modeling of appearance

    The main component of our detector is a statistical representation of visual appearance.  We use a statistical model to represent the object's appearance over a small range of pose variation.  The purpose of this model is to capture variation in the appearance of the object  that cannot be modelled explicitly.  This includes variation in the object itself, variation due to lighting, and small variations in pose.  We also use a statistical model to describe the rest of the visual world; that is, everything the object must be discriminated from.

    2.  Multiple view-specific detectors  - Since each detector is designed for a specific orientation of the object, we use multiple detectors that span a range of the object's orientation.  We then combine the results of these individual detectors:

    3.  Exhaustive search -  The desired object may exist at any size and position within the image.  To detect the object at any position, we must exhaustively apply the detectors over a range of positions.  To detect the object at any size we must also apply the detectors to scaled versions of the image.


    Frontal Face detection

    Test set: 130 images, 507 frontal faces.  Images selected by Sung , Poggio and Rowley, Baluja, Kanade
    Notes Detection rate False detections
    Our detector Excluded 5 images of 
    hand drawn faces
    90.5% 2
    Rowley, Baluja, Kanade 86.0% 31
    Colmenarez, Huang 89.9% 6,1333