Efthymiadis Nikos ; Tolias Giorgos ; Chum Ondrej

Abstract

This work addresses scaling up the sketch classification task into a largenumber of categories. Collecting sketches for training is a slow and tediousprocess that has so far precluded any attempts to large-scale sketchrecognition. We overcome the lack of training sketch data by exploiting labeledcollections of natural images that are easier to obtain. To bridge the domaingap we present a novel augmentation technique that is tailored to the task oflearning sketch recognition from a training set of natural images.Randomization is introduced in the parameters of edge detection and edgeselection. Natural images are translated to a pseudo-novel domain called"randomized Binary Thin Edges" (rBTE), which is used as a training domaininstead of natural images. The ability to scale up is demonstrated by trainingCNN-based sketch recognition of more than 2.5 times larger number of categoriesthan used previously. For this purpose, a dataset of natural images from 874categories is constructed by combining a number of popular computer visiondatasets. The categories are selected to be suitable for sketch recognition. Toestimate the performance, a subset of 393 categories with sketches is alsocollected.