Automated adjustment of crowdsourced calorie estimations for accurate food image logging

Obesity is increasing globally and is a risk factor for many chronic conditions such as such as heart disease, sleep apnea, type-2 diabetes, and some cancers. Research shows that food logging is beneficial in promoting weight loss. Crowdsourcing has also been used in promoting dietary feedback for food logging. This work investigates the feasibility of crowdsourcing to provide support in accurately determining calories in meal images. Two groups, 1. experts and 2. non-experts, completed a calorie estimation survey consisting of 15 meal images. Descriptive statistics were used to analyse the performance of each group. Collectively, non- experts could determine which meals had larger amounts of calories and analysis showed that meals with greater calories resulted in greater standard deviations of non-expert estimates. Secondary experiments were completed that used crowdsourcing to adjust user calorie estimations using non-expert calorie estimations. Five-fold cross validation was used and results from the calorie adjustment process show a reduced overall mean calorie difference in each fold and the mean error percentage decreased from 40.85% to 25.52% in comparing original mean estimations against adjusted mean estimations. As such, there is credibility in adjusting calorie estimates from a crowd as opposed to simply taking a central measure such as the mean.