A study on cost reductions in labeling annotation of food images

Abstract

In this paper, we discuss the cost reduction in labeling annotation of food images. Since the number of categories of food images is large, it takes time to select an appropriate label. We examine whether the cost reduction is possible by answering only the success or failure of the candidate label presented in advance by food image classifier. We compared the labeling speed with the method in which labeling workers watch the food image and select an appropriate label without any candidate labels. As a result, we concluded the cost reduction is possible in some categories. In other categories, we concluded improving the classifier accuracy is needed for the cost reduction.