Making Templates Rotationally Invariant: An Application to Rotated Digit Recognition

[baluja@vie.ius.cs.cmu.edu]

The following paper is available from:
http://www.cs.cmu.edu/~baluja


Making Templates Rotationally Invariant:
An Application to Rotated Digit Recognition

Shumeet Baluja



Abstract:
This paper describes a simple and efficient method to make
template-based object classification invariant to in-plane
rotations. The task is divided into two parts: orientation
discrimination and classification. The key idea is to perform the
orientation discrimination before the classification. This can be
accomplished by hypothesizing, in turn, that the input image belongs
to each class of interest. The image can then be rotated to maximize
its similarity to the training images in each class (these contain
the prototype object in an upright orientation). This process yields
a set of images, at least one of which will have the object in an
upright position. The resulting images can then be classified by
models which have been trained with only upright examples. This
approach has been successfully applied to two real-world vision-based
tasks: rotated handwritten digit recognition and rotated face
detection in cluttered scenes.
 

This work was completed while the author was at:
Justsystem Pittsburgh Research Center &
School of Computer Science, Carnegie Mellon University



Comments and Questions welcome.  
Please send all feedback to sbaluja at lycos.com.