Tufts iDine Dataset

Dietary intake influences disease risk and coresponding economic and healthcare burdens worldwide. In recent years, a variety of tools have emerged for assessing dietary intake, for both clinical use in the healthcare sector and for consumer mobile apps used as part of weight-loss or recovery programs. Deep learning has the ability to greatly reduce this cognitive burden by measuring the volume of foods in a single picture, improving both accuraacy and ease of use of these tools. Few deep learning-based dietary assessment tools exist, due in part to the large amounts of precisely-annotated data required to train these models. To date, no existing food dataset contains annotations for classification, segmentation, depth, and nutritional information. We introduce iDine, a first-of-its-kind food image dataset that includes meals with varied quantities of leftovers as well as comprehensive annotations for detection, classification, and segmentation.