Food vs Non-Food Classification

Automatic understanding of food is an important research challenge. Food recognition engines can provide a valid aid for automatically monitoring the patient's diet and food-intake habits directly from images acquired using mobile or wearable cameras. One of the first challenges in the field is the discrimination between images containing food versus the others. Existing approaches for food vs non-food classification have used both shallow and deep representations, in combination with multi-class or one-class classification approaches. However, they have been generally evaluated using different methodologies and data, making a real comparison of the performances of existing methods unfeasible. In this paper, we consider the most recent classification approaches employed for food vs non-food classification, and compare them on a publicly available dataset. Different deep-learning based representations and classification methods are considered and evaluated.