What auto completion tells us about sketch recognition

Abstract

Auto completion is generally considered to be a difficult problem in sketch recognition as it requires a decision to be made with fewer strokes. Therefore, it is generally assumed that the classification of fully completed object sketches should yield higher accuracy rates. In this paper, we report results from a comprehensive study demonstrating that the first few strokes of an object are more important than the lastly drawn ones. Once the first few critical strokes of a symbol are observed, recognition accuracies reach a plateau and may even decline. This indicates that less is more in sketch recognition. Our results are supported by carefully designed computational experiments using Tirkaz et. al.'s sketch auto completion framework on the dataset of everyday object sketches collected by Eitz et. al..