3D Reconstruction and Volume Estimation of Food using Stereo Vision Techniques

Abstract

It is generally accepted that a healthy diet plays an important role in modern lifestyle and can prevent or reduce the effects of important diseases, such as obesity, diabetes or cardiovascular diseases. Technological advancement and the wide spread of smartphones enable the monitoring and recording of nutritional habits on a daily basis, through mHealth solutions. The most difficult task of mHealth dietary systems for calculating the nutritional composition of food is to estimate its volume. In this study, we present a volume estimation system based on structure from motion smartphone camera, through two-view 3D food reconstruction. The proposed methodology uses stereo vision techniques and requires the input of two food images with a reference card next to the plate, to reconstruct the 3D structure of the food and to estimate its volume. The above approach achieves a mean absolute percentage error from 4.6 - 11.1% per food dish. The systematic collection of a labelled Mediterranean Greek Food images dataset, the MedGRFood, with known food weight allows the evaluation of the proposed methodology.