The segmented UEC Food-100 dataset with benchmark experiment on food detection

Abstract

Automatic food classification systems have several interesting applications ranging from detecting eating habits, to waste food management and advertisement. When a food image has multiple food items, the food detection step is necessary before classification. This work challenges the food detection issue and it introduces to the research community the Segmented UEC Food-100 dataset, which expands the original UEC Food-100 database with segmentation masks. In the semantic segmentation experiment, the performance of YOLAC and DeeplabV3+ has been compared and YOLAC reached the best accuracy of 64.63% mIoU. In the instance segmentation experiment, YOLACT has been used due to its speed and high accuracy. The benchmark performance on the newly released Segmented UEC Food-100 dataset is 68.83% mAP. For comparison purpose, experiments have been run also on the UEC FoodPix Complete dataset of Okamoto et al. The database and the code will be available after publication.