Diabetes60 — Inferring Bread Units From Food Images Using Fully Convolutional Neural Networks

In this paper we propose a challenging new computer vision task of inferring Bread Units (BUs) from food images. Assessing nutritional information and nutrient volume from a meal is an important task for diabetes patients. At the moment, diabetes patients learn the assessment of BUs on a scale of one to ten, by learning correspondence of BU and meals from textbooks. We introduce a large scale data set of around 9k different RGB-D images of 60 western dishes acquired using a Microsoft Kinect v2 sensor. We recruited 20 diabetes patients to give expert assessments of BU values to each dish based on several images. For this task, we set a challenging baseline using state-of-the-art CNNs and evaluated it against the performance of human annotators. In our work we present a CNN architecture to infer the depth from RGB-only food images to be used in BU regression such that the pipeline can operate on RGB data only and compare its performance to RGB-D input data. We show that our inferred depth maps from RGB images can replace RGB-D input data at high significance for the BU regression task. In its best configuration, our proposed method achieves a RMSE of 1.53 BUs using RGB and inferred depth. Considering the variability among the raters themselves of RMSE = 0.89, we can show that our baseline method with depth prediction can extract reasonable nutritional information from RGB image data only.