Perceptually Based Learning of Shape Descriptions for Sketch Recognition

by

Olya Veselova and Randall Davis

Summary

A huge chunk of this paper can be summarized through one quote, "people pay unequal attention to different features". The goal for the authors is to have a system be able to learn descriptions from a single example. The system should only capture the relevant features that users would care about. The paper made heavy use of Goldmeier's work on human perception. Goldmeier had properties called singularities, which were features that small variations in had a qualitative difference. Goldmeier's singularities include vertically, symmetry, parallelism, horizontally and straightness. The paper lists the importance of these different constraints. This allows for a score. The score can be adjusted by obstruction, tension lines, and grouping. An example is if two lines are near they being parallel is an important constraint. If they are far away and a number of primitives are between them that constraint isn't so important.

To test their system the created a study and measured how often their system agreed with people's perceptual judgments on near-perfect drawings.

Discussion

The paper reminds me of those standardized test in which they student is given multiple shapes in a group and must recognize which shape does not belong. The work is interesting and has a direct connection to how we make our domain descriptions in LADDER. We should only focus on the constraints that the user believes to be relevant.

Citation

Veselova, O. and Davis, R. 2006. Perceptually based learning of shape descriptions for sketch recognition. In ACM SIGGRAPH 2006 Courses (Boston, Massachusetts, July 30 - August 03, 2006). SIGGRAPH '06. ACM Press, New York, NY, 28. DOI= http://doi.acm.org/10.1145/1185657.1185789