Fig. 4: QuickDraw only-one-shot doodle classification.

a shows one sample run that illustrates the training images, test images, and a unit task marked in red. b shows test accuracies from various models, including ours and human performance. Accuracy is reported as a percentage, with error bars indicating the mean and the standard deviation across five independent runs. Error bars are omitted for deterministic models. c displays inter-rater agreement, measured by Fleiss’ kappa, between our model and human performances. It shows that our model not only achieves near-human accuracy but also makes similar mistakes. Human results are from five healthy subjects, H₁–H₅, aged 20 to 29.