Similarity is among the most basic concepts studied in psychology. Yet, there is no unique way of assessing similarity of two objects. In the sketch recognition domain, many tasks such as classification, detection or clustering require measuring the level of similarity between sketches. In this paper, we propose a carefully designed experiment setup to construct a gold standard for measuring the similarity of sketches. Our setup is based on table scaling, and allows efficient construction of a measure of similarity for large datasets containing hundreds of sketches in reasonable time scales. We report the results of an experiment involving a total of 9 unique assessors, and 8 groups of sketches, each containing 300 drawings. The results show high interrater agreement between the assessors, which makes the constructed gold standard trustworthy.
Authors: Serike Cakmak, T. Metin Sezgin.