Investigation of Methodologies of Food Volume Estimation and Dataset for Image-Based Dietary Assessment

Abstract

In recent years, lifestyle has changed massively due to socio-economic effect. These changes in lifestyle have unfortunately been leading to increase in chronic illnesses. With flourishing digital market of various communication devices, a variety of tools which assist health-lovers to track their lifestyle with eating habits, exercising patterns and calorie intake have come up. Most of these tools provide image-based automated dietary assessment. Image-based automated methods to estimate volume of food image present accurate methods to estimate calorie count. At the background of these tools are complex conventional image-based and the latest deep learning algorithms. In this work, we present a review of these methods. An evaluation of these algorithms show that conventional methods like 3D reconstruction method gives accuracy of about 94.4% and deep learning approach gives accuracy of 95.41% and outperform other methods. Another important feature that contributes to the accuracy of the tool is the training image dataset. The study shows that a different variety of datasets were used for these methods with 3D reconstruction method giving error in the range of 0.51–17.55% and deep learning in the range of 1.62–9.28% for different food items. The datasets used comprise of only raw foods, tinned foods, and packaged foods. Thereby, the work presents the research gap for benchmark dataset that addresses diverse food types and produces uniform training sample set for the estimator models to compare their relative performance in food volume estimation methods for different food types.