Performance Evaluation Methods of Computer Vision Systems for Meal Assessment

Abstract

Several systems have been proposed for the automatic food intake assessment and dietary support by analyzing meal images captured by smartphones. A typical system consists of computational stages that detect/segment the existing foods, recognize each of them, compute their volume, and finally estimate the corresponding nutritional information. Although this newborn field has made remarkable progress over the last years, the lack of standardized datasets and established evaluation frameworks has made difficult the comparison between methods and eventually prevented the formal definition of the problem. In this paper, we present an overview of the datasets and protocols used for evaluating the computer vision stages of the proposed automatic meal assessment systems.