hide
Free keywords:
-
Abstract:
We present an iterative optimization scheme for obtaining minimal line sketches of object categories. Minimal sketches are introduced as a tool to derive the most important visual properties of a visual object category and can potentially provide useful constraints for automatic classification algorithms. We define the minimal sketch of an object category as the minimal number of straight lines necessary to lead to a correct recognition by 75 of naïve participants after one second of presentation. Nine participants produced sketches of 30 object categories. We displayed the three sketches with the lowest number of lines for each category to 24 participants who freely named them. In consecutive rounds the sketchers had to optimize their drawings independently based on sketches and responses of the previous rounds. The optimized sketches were subsequently rated again by 24 new subjects. The average number of lines used in the sketches decreased from 8.8 to 7.9 between the two trials while the average recognition rate increased from 57.3 to 67.9. 27 of the 30 categories had at least one sketch that was recognized by more than 75 of subjects. For most of the categories, the sketches converged to an optimum within two drawing-rating rounds.