Maintenance Notice

Due to necessary scheduled maintenance, the JMIR Publications website will be unavailable from Wednesday, July 01, 2020 at 8:00 PM to 10:00 PM EST. We apologize in advance for any inconvenience this may cause you.

Who will be affected?

Advertisement

Citing this Article

Right click to copy or hit: ctrl+c (cmd+c on mac)

Published on 25.03.20 in Vol 8, No 3 (2020): March

Preprints (earlier versions) of this paper are available at http://preprints.jmir.org/preprint/15294, first published Jun 28, 2019.

This paper is in the following e-collection/theme issue:

    Original Paper

    Volumetric Food Quantification Using Computer Vision on a Depth-Sensing Smartphone: Preclinical Study

    1Department of Diabetes, Endocrinology, Nutritional Medicine and Metabolism, Bern University Hospital, University of Bern, Bern, Switzerland

    2Laboratory of Biometry, University of Thessaly, Nea Ionia/Volos, Magnesia, Greece

    3University Institute of Clinical Chemistry, Inselspital, Bern University Hospital, University of Bern, Bern, Switzerland

    4Insel Data Science Center & Department for Technology and Innovation, Bern University Hospital, University of Bern, Bern, Switzerland

    *these authors contributed equally

    Corresponding Author:

    Lia Bally, MD, PhD

    Department of Diabetes, Endocrinology, Nutritional Medicine and Metabolism

    Bern University Hospital

    University of Bern

    Freiburgstrasse 15

    Bern, 3010

    Switzerland

    Phone: 41 31 632 36 77

    Email: lia.bally@insel.ch


    ABSTRACT

    Background: Quantification of dietary intake is key to the prevention and management of numerous metabolic disorders. Conventional approaches are challenging, laborious, and lack accuracy. The recent advent of depth-sensing smartphones in conjunction with computer vision could facilitate reliable quantification of food intake.

    Objective: The objective of this study was to evaluate the accuracy of a novel smartphone app combining depth-sensing hardware with computer vision to quantify meal macronutrient content using volumetry.

    Methods: The app ran on a smartphone with a built-in depth sensor applying structured light (iPhone X). The app estimated weight, macronutrient (carbohydrate, protein, fat), and energy content of 48 randomly chosen meals (breakfasts, cooked meals, snacks) encompassing 128 food items. The reference weight was generated by weighing individual food items using a precision scale. The study endpoints were (1) error of estimated meal weight, (2) error of estimated meal macronutrient content and energy content, (3) segmentation performance, and (4) processing time.

    Results: In both absolute and relative terms, the mean (SD) absolute errors of the app’s estimates were 35.1 g (42.8 g; relative absolute error: 14.0% [12.2%]) for weight; 5.5 g (5.1 g; relative absolute error: 14.8% [10.9%]) for carbohydrate content; 1.3 g (1.7 g; relative absolute error: 12.3% [12.8%]) for fat content; 2.4 g (5.6 g; relative absolute error: 13.0% [13.8%]) for protein content; and 41.2 kcal (42.5 kcal; relative absolute error: 12.7% [10.8%]) for energy content. Although estimation accuracy was not affected by the viewing angle, the type of meal mattered, with slightly worse performance for cooked meals than for breakfasts and snacks. Segmentation adjustment was required for 7 of the 128 items. Mean (SD) processing time across all meals was 22.9 seconds (8.6 seconds).

    Conclusions: This study evaluated the accuracy of a novel smartphone app with an integrated depth-sensing camera and found highly accurate volume estimation across a broad range of food items. In addition, the system demonstrated high segmentation performance and low processing time, highlighting its usability.

    JMIR Mhealth Uhealth 2020;8(3):e15294

    doi:10.2196/15294

    KEYWORDS



    Introduction

    Qualitative and quantitative assessment of dietary intake are cornerstones for the prevention and management of metabolic diseases such as obesity and diabetes [1,2]. Traditional manual food records that rely on human abilities to quantify food intake are time-consuming and error-prone [3]. One of the main challenges is the appropriate estimation of portion size (ie, volume) [4]. Inaccurate portion size estimation contributes up to 50% of the total estimation error [5]. Novel approaches replacing manual input by automated techniques may overcome the inherent limitations of traditional approaches, while increasing usability.

    Mobile devices, currently ubiquitous, could simplify dietary monitoring. Although there are a number of commercially available apps offering access to food composition databases or providing reference images to facilitate estimation of portion size [6], they are generally limited by the need for manual user input.

    High-quality smartphone cameras and computer vision approaches can be combined to fully automate portion size estimation. Users capture images of the meal using the smartphone camera, and the app subsequently builds a 3D model of the food to calculate its volume [7]. Combining the food volume with macronutrient-density databases, the app translates the volume into weight and then nutrient information. Food identification can be accomplished either by user selection or as part of the automated image processing, which further minimizes the need for user input [8].

    Researchers have described several such systems [9-13]. A major challenge in many of these approaches lies in the capturing of the third dimension (depth) due to geometric constraints. In particular, factors such as precise food location, shape and size of food items, and changes in these parameters depending on camera perspective potentially interfere with reliable depth assessment. To overcome such constraints, fiducial markers, which ground the scene in a common frame of reference, are utilized. In addition, some systems use multiple images or video sequences of the food, followed by a complex calibration process. All these aspects inherently affect usability and accuracy.

    The recent advent of miniaturized depth-sensing cameras embedded within smartphones (eg, iPhone X) opens a new horizon for automated food quantification. Using a single capture including depth information from any convenient viewing angle, this technology has the potential to eliminate the need for manual input, thereby increasing usability as well as accuracy. Therefore, the aim of this study was to evaluate the accuracy of a novel smartphone app that combines depth-sensing with computer vision to quantify food volume across a broad range of meals reflecting a real-life setting.


    Methods

    Study Design

    The study occurred at the Central Kitchen Facility of the University Hospital Bern, Switzerland, in mid-January 2019. The system was tested on regular meals served to patients and hospital staff. A total of 48 test meals were randomly generated from a pre-defined pool of 128 food items. The test meals comprised the following meal types: breakfast, cooked meals, and snacks. Meals consisted of 4 (breakfast), 3 (cooked meals), or 1 (snacks) food items and were served on a standard plate or in a standard bowl. The meal content is described in Multimedia Appendix 1.

    For each test meal, a single image was taken at a predefined angle of either 45° or 90° from the horizontal position (the angle was estimated by the user). The allocation of the capture angle of each meal was pre-defined using a balanced randomization procedure. The randomization sequence was produced as a binary sequence in three batches (by meal type). Images were captured under natural light conditions.

    Smartphone App

    The app was provided by SNAQ GmbH (Zurich, Switzerland) using a software version from May 2018, and it was installed on an iPhone X (Apple Inc, Cupertino, California), which uses a built-in depth sensor applying structured light. The automated food quantification workflow is summarized in Figure 1 and consisted of (1) capturing the scene, (2) analyzing the scene, (3) estimating the volume, (4) converting the food volume into food weight, and (5) conversion of the food weight into macronutrient content.

    Figure 1. Automated food quantification workflow performed by the SNAQ app.
    View this figure

    First, the user takes a photograph using the phone, and a depth map of the food is generated through the phone’s front sensors consisting of a photo camera and an active depth sensor.

    Second, the system partitions the image into consistent regions representing different items and eliminates those that are not food. To do so, a convolutional neural network has learned how food is structured in terms of sets of pixels and their correlations to the visual appearance of images. The data used to train the system consist of images with flags for each pixel indicating whether the respective pixel represents food. If the automated segmentation is not deemed satisfactory, the user can manually adjust the outlines of the items. A workflow of the segmentation as well as an example of good and bad segmentations are provided in Multimedia Appendix 2. Then, based on the depth map and input from the phone sensors, the visible point cloud is transformed into a set of surfaces using a Delaunay triangulation. The system extracts the location and orientation of the table (vertical plane) using the RANSAC-Algorithm [14] for an outlier-robust fitting. From this, the surfaces of each dish are defined. Selection of the food type is manually performed for each of the segmented food items.

    Third, the segmented food items are used to cut the visible surface into partial food surfaces. Each food surface is then closed by the dish surfaces before their volume is calculated.

    Fourth, the food volume is converted into food weight using a food density database. Finally, the food weight is converted into macronutrient content using the Swiss Food Composition Database [15].

    This study was designed to assess the accuracy of automated quantification of portion sizes. Automated food recognition (ie, the taxonomy of the food items) was not a focus of this study. Instead, the user capturing the image selected the respective food item from a pre-defined list within the app.

    Reference Method

    The reference weight was generated by weighing individual food items to the nearest 0.1 g using a precision scale (ME4002, Mettler Toledo, Greifensee, Switzerland). Conversion into macronutrient content was performed using the Swiss Food Composition Database [15].

    Endpoints

    There were four study endpoints: (1) error of the estimated meal weight, (2) error of the estimated meal macronutrient content and energy content, (3) segmentation performance (defined as the number of items requiring manual correction of segmentation as well as intersection of the uncorrected and corrected segmentation areas over the corrected segmentation area), and (4) processing time (defined as the time period from image capture to macronutrient/energy output, including the time required for manual inputs).

    Sample Size Calculation and Statistical Analysis

    The number of test meals in this study was determined based on a pilot experiment showing a mean (SD) difference in carbohydrate content of –2.6 g (9.2 g). Applying a power of 80% and significance level of .05 for 48 meals was deemed appropriate.

    The error was determined on the meal level, and the following error metrics were used: bias, defined as the difference between estimation and reference (estimate-reference); absolute error, defined as |estimate-reference|; and 95% limits of agreement, calculated as ±2*SD of the bias. Bland-Altman plots were generated to visualize the level of agreement between the estimate and reference values. General linear models were used to assess the effect of meal type and inclination angle on the estimation error. P values <.05 were considered statistically significant. SPSS version 25.0 (IBM Corp, Armonk, NY) was used for statistical analysis. Data are described using mean (SD) and median (interquartile range [IQR]). All absolute error and bias values in this paper are presented as absolute values (g) followed by the relative values (%) in parentheses.


    Results

    Macronutrient and Energy Content of the Test Meals

    The 48 test meals encompassed 128 food items. The mean reference macronutrient and energy contents of the 48 test meals are summarized in Multimedia Appendix 3. On average, the meals weighed 235.8 g (range 29.6-582.4 g). Meals contained an average 38.5 g carbohydrate (range 4.4-101.0 g), 14.6g protein (range 0.2-66.9 g), and 11.7 g fat (range 0.1-37.1 g). Mean energy content was 325 kcal (range 32-609 kcal). Insights into the study meals, including the representation of different meal types, are provided in Multimedia Appendix 4.

    Errors of Estimated Meal Weight, Estimated Meal Macronutrient Content, and Estimated Meal Energy Content

    The mean (SD) error metrics are summarized in Table 1, and the median (IQR) error metrics are presented in Multimedia Appendix 5. Corresponding Bland-Altman plots are presented in Figures 2-6. In both absolute and relative terms, the mean (SD) absolute error of the estimated weight for all meals was 35.1 g (42.8 g; 14.0% [12.2%]), and the mean (SD) bias was 19.3 g (52.1 g; 5.4% [17.8%]). The 95% limits of agreement were –84.8 g and 123.4 g (Figure 2).

    Table 1. Error metrics for all meals and by meal, reported as the difference between the app’s estimate and the reference weight, macronutrient content, or energy content.
    View this table
    Figure 2. Bland-Altman plot illustrating the difference between the estimated and reference meal weights.
    View this figure

    The mean (SD) absolute error of the estimated carbohydrate content for all meals was 5.5 g (5.1 g; 14.8% [10.9%]), and the mean (SD) bias was 1.0 g (7.5 g; 2.9% [18.3%]). The 95% limits of agreement were –13.9 g and 15.9 g (Figure 3). The mean (SD) absolute error of the estimated protein content for all meals was 2.4 g (5.6 g; 13.0% [13.8%]), and the mean (SD) bias was 1.7 g (5.9 g; 5.6% [18.2%]). The 95% limits of agreement were –10.0 g and 13.4 g (Figure 4). The mean (SD) absolute error of the estimated fat content for all meals was 1.3 g (1.7 g; 12.3% [12.8%]), and the mean (SD) bias was 0.5 g (2.1 g; 5.7% [16.9%]). The 95% limits of agreement were –3.8 g and 4.7 g (Figure 5). The mean (SD) absolute error of the estimated energy content for all meals was 41.2 kcal (42.5 kcal; 12.7% [10.8%]), and the mean (SD) bias was 15.5 kcal (57.4 kcal; 4.1% [16.2%]). The 95% limits of agreement were –99.4 kcal and 130.3 kcal (Figure 6).

    Figure 3. Bland-Altman plot illustrating the difference between the estimated and reference carbohydrate content of the meals.
    View this figure
    Figure 4. Bland-Altman plot illustrating the difference between the estimated and reference protein content of the meals.
    View this figure
    Figure 5. Bland-Altman plot illustrating the difference between the estimated and reference fat content of the meals.
    View this figure
    Figure 6. Bland-Altman plot illustrating the difference between the estimated and reference energy content of the meals.
    View this figure

    While viewing angle had no significant influence on the accuracy of macronutrient and energy estimation (P=.96, P=.83, P=.99, P=.73, and P=.70 for absolute errors in weight, carbohydrate, protein, fat, and energy content, respectively), we observed a significant effect of meal type on the accuracy for all macronutrients and energy content (P=.001, P=.001, P=.001, P=.002, and P=.005 for absolute errors in weight, carbohydrate, protein, fat, and energy content, respectively). The mean bias values of the cooked meals for carbohydrate, protein, and fat content were significantly higher than for the breakfasts, with a mean (SD) difference in bias of 18.3% (4.9%) for carbohydrate, 19.6% (4.9%) for protein, and 17.7% (5.1%) for fat (all P<.001). The comparison of bias for snacks relative to the bias of cooked meals resulted in marginal outcomes for fat (P=.02), carbohydrate (P=.08), and protein (P=.07). Furthermore, the bias for snacks was not different from the bias for breakfasts for all macronutrients (P=.37, P=.26, P=.23, P=.79, and P=.50 for weight, carbohydrate, protein, fat, and energy content, respectively).

    Segmentation Performance

    In 7 of the 128 items (5.5%), segmentation required manual adjustment. The intersection over union of unadjusted to adjusted segmentation area was 71.8%.

    Processing Time

    Mean (SD) processing time across all meal types was 22.9 seconds (8.6 seconds). Processing time was significantly lower for snacks (mean 17.9 seconds, SD 7.0 seconds) compared with cooked meals (mean 27.8 seconds, SD 10.8 seconds; mean difference –9.9 seconds, SD 2.7 seconds; P<.001). Processing time was lower for breakfast (mean 23.1 seconds, SD 3.5 seconds) compared with cooked meals (mean difference –4.7 seconds, SD 2.8 seconds; P=.12). Figure 7 provides the processing time stratified by meal type.

    Figure 7. Box-plot of the processing time according to meal type. Box-plots show median values (solid line), interquartile range (IQR; box outline), spread of data points without outliers (whiskers) and outliers identified as 1.5*IQR (symbols).
    View this figure

    Discussion

    This study evaluated the accuracy of a novel smartphone app that combines depth sensing with computer vision using volumetry to quantify the macronutrient content of meals in a real-life setting. The main findings were as follows: the accuracy was adequate across all macronutrients, the accuracy differed according to meal type (lower for cooked meals than for snacks and breakfast), segmentation was good, and processing was fast.

    When compared with previous reports of apps using computer vision without depth sensors, the present app had comparable, or even superior, accuracy. Rhyner et al [16] reported a mean absolute error of 26.2% in carbohydrate content when assessing 60 cooked meals with non-overlapping food items. In a further preclinical study assessing the accuracy of the prototype used by Rhyner et al and based on 54 cooked meals, the mean absolute error in carbohydrate quantification was 14.8 g, which corresponds to 24.7% for a meal carbohydrate content of 60 g [17]. In contract, with the app in the present study, mean absolute errors in macronutrient content estimation ranged from 12.3% (fat) to 15% (carbohydrate).

    Of note, two recent studies assessing the accuracy of image–based food quantification using volume as a reference metric reported mean absolute errors in volume estimation of 7.2% [12] and 5.8% [18] based on the assessment of 5 and 20 food items, respectively. These slightly smaller errors compared to those in this study can be explained by the different reference metric used to define the system accuracy (error in estimated volume versus error in estimated weight and consequently macronutrient content). Of note, errors in weight estimation have two potential sources: inaccuracies in volume and density estimation. Additionally, operational aspects of the previously reported systems differ from those in this study. Xu et al [12] used a complex multi-step approach including reference objects, while Makhsous et al [13] added a depth sensor with structured light to the smartphone and complemented their approach with video sequences, significantly increasing the complexity of the workup. These differences highlight the important tradeoffs between accuracy and usability.

    Of note, this study revealed a comparably short processing time, ranging from 18 seconds for snacks to 29 seconds for cooked meals. This is faster than those reported in previous studies, where processing times generally exceeded the limit of 1 minute [19]. This highlights the usability of the present system even when applied to meals in a real-life setting.

    The accuracy of the tested app differed according to meal type and was lower for cooked meals than for breakfasts and snacks. This might have resulted from the different levels of complexity in terms of scene analysis of the respective meals. Whereas the breakfasts and snacks had food items that were clearly separated from each other, the cooked meals had food items with touching borders or a certain degree of overlap. Notably, the angle of image capture did not affect the estimation accuracy in this study, indicating the flexibility, usability, and robustness of the system.

    We acknowledge a number of limitations of this study. First, the assessment was limited to meals provided by the hospital kitchen, preventing a generalized statement on the accuracy. However, the system was tested using real-life meals, underscoring its potential use in practice. Second, the system was limited to a single type of smartphone (iPhone X) with a depth sensor, precluding statements on the performance of the software combined with different hardware components. However, this approach supports the strength of providing a commercially available tool. Third, the depth sensor limited the reconstruction to 1/20th of the resolution and with lower depth precision than with a passive depth sensor (dual camera approach). However, the use of a depth sensor foregoes the need for fiducial markers, rendering it more convenient to users. Fourth, we served all meals on one plate or bowl type, possibly reducing the variation in volume estimation that was unrelated to the depth sensor. Finally, this study exclusively focused on the accuracy of volume quantification and did not consider food recognition.

    When considering both observed accuracy and usability of the present system, the field of potential use appears broad. Such a system may be of interest in the medical sector to assist with nutritional counseling and management of patients with metabolic disorders (eg, diabetes mellitus, obesity) or at risk of malnutrition. Beyond this, such a system may be valuable in nutritional epidemiology due to the potential to systematically and accurately monitor dietary intake on a large scale.

    In conclusion, this study evaluated the accuracy of a novel smartphone app with integrated depth sensing and found a high level of accuracy in volumetric macronutrient and energy estimation across a broad set of meals in a real-life setting. In addition, the system demonstrated high segmentation performance and low processing time, highlighting its usability.

    Acknowledgments

    We thank the staff from the Bern University Hospital Kitchen (Thomas Walser, Beat Blum, Vinzenz Meier) for provision of food items and tableware. We also extend special thanks to Michèle Monnard for support with data management and to the developer of the software SNAQ (Aurelian Briner, Nico Previtali, Lukas Frischknecht, and the team) for the pleasant collaboration. This project was supported by the Diabetes Center Bern and the Berner Kantonalbank (BEKB).

    Authors' Contributions

    DH, CN, CS, and LB conceived and designed the study. DH, CK, CL, and RJ acquired the data. DH, CN, and LB analyzed the data. DH, CN, JD, and LB interpreted the data. All authors were involved in drafting or revising the article for important intellectual content. All authors approved the final version of the manuscript. CS and LB are responsible for the integrity of the work as a whole.

    Conflicts of Interest

    None declared.

    Multimedia Appendix 1

    Food items assessed using the SNAQ smartphone app.

    DOCX File , 15 KB

    Multimedia Appendix 2

    Examples of good and bad segmentation by the SNAQ app during test meal analysis.

    PNG File , 1766 KB

    Multimedia Appendix 3

    Mean macronutrient and energy content of the tested meals.

    DOCX File , 16 KB

    Multimedia Appendix 4

    Examples of the assessed meal types.

    PNG File , 1301 KB

    Multimedia Appendix 5

    Estimation accuracy, as determined using the error metrics of the analyzed meals.

    DOCX File , 15 KB

    References

    1. American Diabetes Association. Evidence-based nutrition principles and recommendations for the treatment and prevention of diabetes and related complications. Diabetes Care 2002 Jan;25(1):202-212. [CrossRef] [Medline]
    2. Raynor HA, Champagne CM. Position of the Academy of Nutrition and Dietetics: Interventions for the Treatment of Overweight and Obesity in Adults. J Acad Nutr Diet 2016 Jan;116(1):129-147. [CrossRef] [Medline]
    3. Livingstone MB, Prentice AM, Strain JJ, Coward WA, Black AE, Barker ME, et al. Accuracy of weighed dietary records in studies of diet and health. BMJ 1990 Mar 17;300(6726):708-712 [FREE Full text] [CrossRef] [Medline]
    4. Ordabayeva N, Chandon P. In the eye of the beholder: Visual biases in package and portion size perceptions. Appetite 2016 Aug 01;103:450-457. [CrossRef] [Medline]
    5. Beasley J, Riley WT, Jean-Mary J. Accuracy of a PDA-based dietary assessment program. Nutrition 2005 Jun;21(6):672-677. [CrossRef] [Medline]
    6. Maringer M, Van't Veer P, Klepacz N, Verain MCD, Normann A, Ekman S, et al. User-documented food consumption data from publicly available apps: an analysis of opportunities and challenges for nutrition research. Nutr J 2018 Jun 09;17(1):59 [FREE Full text] [CrossRef] [Medline]
    7. Chowdhary CL. 3D Object Recognition System Based On Local Shape Descriptors and Depth Data Analysis. CSENG 2019 Jan 10;12(1):18-24. [CrossRef]
    8. Ciocca G, Napoletano P, Schettini R. Food Recognition: A New Dataset, Experiments, and Results. IEEE J. Biomed. Health Inform 2017 May;21(3):588-598. [CrossRef]
    9. Anthimopoulos M, Dehais J, Shevchik S, Ransford BH, Duke D, Diem P, et al. Computer vision-based carbohydrate estimation for type 1 patients with diabetes using smartphones. J Diabetes Sci Technol 2015 May;9(3):507-515 [FREE Full text] [CrossRef] [Medline]
    10. He H, Kong F, Tan J. DietCam: Multiview Food Recognition Using a Multikernel SVM. IEEE J. Biomed. Health Inform 2016 May;20(3):848-855. [CrossRef]
    11. Chen H, Jia W, Yue Y, Li Z, Sun Y, Fernstrom JD, et al. Model-based measurement of food portion size for image-based dietary assessment using 3D/2D registration. Meas Sci Technol 2013 Oct;24(10) [FREE Full text] [CrossRef] [Medline]
    12. Xu C, He Y, Khanna N, Boushey C, Delp E. Model-Based Food Volume Estimation Using 3D Pose. In: Proc Int Conf Image Proc. 2013 Sep Presented at: International Conference on Image Processing; 2013; Melbourne, Australia p. 2534-2538   URL: http://europepmc.org/abstract/MED/28572746 [CrossRef]
    13. Makhsous S, Mohammad HM, Schenk JM, Mamishev AV, Kristal AR. A Novel Mobile Structured Light System in Food 3D Reconstruction and Volume Estimation. Sensors (Basel) 2019 Jan 29;19(3) [FREE Full text] [CrossRef] [Medline]
    14. Derpanis K. Overview of the RANSAC Algorithm. Computer Science, York University, Tech. Rep 2010.
    15. Heise T, Meiffren G, Alluis B, Seroussi C, Ranson A, Arrubla J, et al. BioChaperone Lispro versus faster aspart and insulin aspart in patients with type 1 diabetes using continuous subcutaneous insulin infusion: A randomized euglycemic clamp study. Diabetes Obes Metab 2018 Dec 18. [CrossRef] [Medline]
    16. Rhyner D, Loher H, Dehais J, Anthimopoulos M, Shevchik S, Botwey RH, et al. Carbohydrate Estimation by a Mobile Phone-Based System Versus Self-Estimations of Individuals With Type 1 Diabetes Mellitus: A Comparative Study. J Med Internet Res 2016 May 11;18(5):e101 [FREE Full text] [CrossRef] [Medline]
    17. Vasiloglou M, Mougiakakou S, Aubry E, Bokelmann A, Fricker R, Gomes F, et al. A Comparative Study on Carbohydrate Estimation: GoCARB vs. Dietitians. Nutrients 2018 Jun 07;10(6) [FREE Full text] [CrossRef] [Medline]
    18. Puri MZ, Yu Q, Divakaran A, Sawhney H. Recognition and volume estimation of food intake using a mobile device. In: IEEE.: Workshop on Application of Computer Vision; 2009 Presented at: Workshop on Applications of Computer Vision (WACV); 7-8 December, 2009; Snowbird, UT. [CrossRef]
    19. Eldridge A, Piernas C, Illner A, Gibney M, Gurinović M, de Vries J, et al. Evaluation of New Technology-Based Tools for Dietary Intake Assessment-An ILSI Europe Dietary Intake and Exposure Task Force Evaluation. Nutrients 2018 Dec 28;11(1) [FREE Full text] [CrossRef] [Medline]


    Abbreviations

    IQR: interquartile range.


    Edited by G Eysenbach; submitted 28.06.19; peer-reviewed by F Zhu, N Alves, S Guness; comments to author 14.10.19; revised version received 05.11.19; accepted 16.12.19; published 25.03.20

    ©David Herzig, Christos T Nakas, Janine Stalder, Christophe Kosinski, Céline Laesser, Joachim Dehais, Raphael Jaeggi, Alexander Benedikt Leichtle, Fried-Michael Dahlweid, Christoph Stettler, Lia Bally. Originally published in JMIR mHealth and uHealth (http://mhealth.jmir.org), 25.03.2020.

    This is an open-access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work, first published in JMIR mHealth and uHealth, is properly cited. The complete bibliographic information, a link to the original publication on http://mhealth.jmir.org/, as well as this copyright and license information must be included.