IJCAI07-Final - Incremental Learning of Perceptual...

Info iconThis preview shows pages 1–2. Sign up to view the full content.

View Full Document Right Arrow Icon
Abstract Most existing sketch understanding systems require a closed domain to achieve recognition. This paper de- scribes an incremental learning technique for open- domain recognition. Our system builds generalizations for categories of objects based upon previous sketches of those objects and uses those generalizations to clas- sify new sketches. We represent sketches qualitatively because we believe qualitative information provides a level of description that abstracts away details that dis- tract from classification, such as exact dimen- sions. Bayesian reasoning is used in building represen- tations to deal with the inherent uncertainty in percep- tion. Qualitative representations are compared using SME, a computational model of analogy and similarity that is supported by psychological evidence, including studies of perceptual similarity. We use SEQL to pro- duce generalizations based on the common structure found by SME in different sketches of the same ob- ject. We report on the results of testing the system on a corpus of sketches of everyday objects, drawn by ten different people. 1 Introduction The problem of sketch recognition has received much atten- tion in recent years because sketching provides a convenient and natural interface for transferring information from a person to a computer. This problem can be extremely diffi- cult because everyone sketches differently and a single per- son will often sketch the same thing in a different way each time. The key is to identify the properties that remain con- stant across each sketch of a given object. In order to deal with this quandary, many programs use a narrow domain containing a small set of possible sketch objects [e.g., circuit diagrams: Liwicki and Knipping, 2005; simple symbols: Anderson et al ., 2004; architectural objects: Park and Kwon, 2003]. Thus, the programmers can examine the domain ahead of time and either hand-code the classifiers them- selves or train the classifiers on a large body of data (700 images for Liwicki and Knipping [2005]). Even systems designed to work in multiple domains require a certain amount of preprogramming for each particular domain [Al- varado et al., 2002]. While these types of systems have certainly proven useful, they limit the communication be- tween the person and the computer. Only information based in domains that the programmers expect the system to work in can be transmitted. We believe the key to recognition in the absence of domain expectations is efficient, on-line learning. This means that while a user works with the system, it should be learning from the sketches the user produces, so that when the user sketches an object that has been sketched in the past, it will recognize that object. Such a system has a cou- ple of key requirements. Firstly, there must be a simple way for the user to tell the system what a sketched object is sup- posed to be. Secondly, an algorithm that can learn a new category based on only a few examples is required. This is
Background image of page 1

Info iconThis preview has intentionally blurred sections. Sign up to view the full version.

View Full DocumentRight Arrow Icon
Image of page 2
This is the end of the preview. Sign up to access the rest of the document.

This note was uploaded on 06/13/2011 for the course CAP 6938 taught by Professor Staff during the Spring '08 term at University of Central Florida.

Page1 / 6

IJCAI07-Final - Incremental Learning of Perceptual...

This preview shows document pages 1 - 2. Sign up to view the full document.

View Full Document Right Arrow Icon
Ask a homework question - tutors are online