Perception

Perceptually Based Learning of Shape Descriptions for Sketch Recognition - Olya Veselova and Randall Davis

Veselova seeks to learn to recognize sketches based on ideas derived from human perception. After low level recognition, higher level shapes are constructed based on constraints. These constraints are based on how people perceive similar shapes. People base similarity on several "singularities" determined by previously by Goldmeier. They include verticality, horizontality, straightness, parallelism, and others. These properties are translated into constraints. In addition, humans place a greater emphasis on some of these singularities. Thus the constraints are ranked according to priority. Priorities are adjusted depending on bases on three global properties: obstruction, tension lines, and grouping. Obstruction depends on the amount of stuff between two constrained objects; the more stuff the lower the constraint is ranked. Tension lines form a "balance" in the shape and is oriented along a grid system. Grouping is determined in two ways: connectedness and subshapes (similar to shapes making more complex shapes in LADDER). Finally, when compared to user determined similarity, the system achieved a 77% compatibility with the majority vote of similarity. The system performed even better on shapes that users were in high agreement.

Discussion: As noted in the final section of the paper the constraints are grouped into broad classes leaving a good deal of wiggle room for similarity. Also, the authors note that the system lacks global symmetry constraints, a key singularity described earlier in the paper. Such a constraint would seem to be one the more important constraints added. Also lacking is a negation constraint; however, this is a problem in many constraint systems.