Retrieval and Classification of Food Images

G. M. Farinella, D. Allegra, M. Moltisanti, F. Stanco, S. Battiato

Computers in Biology and Medicine, vol. 77, Pages 23–39, doi:10.1016/j.compbiomed.2016.07.006, 2016

  Download the Paper

CONTACT:
Giovanni Maria Farinella
gfarinella@dmi.unict.it


We investigate the problem of food understanding and present a survey about the state of the art. In this page you can find the files and the information used to perform the experiments presented in the paper.


UNICT Food Dataset 1200
The UNICT-FD1200 dataset is composed by 4754 images related to 1200 distinct dishes of food of different nationalities (e.g., English, Japanese, Indian, Italian, Thai,etc.). Each plate has been acquired multiple times (four in the average) to guarantee the presence of geometric and photometric variabilities. All the food photos have been taken in the last ve years during real meals by using a mobile camera in unconstrained settings, such as diefferent backgrounds and light conditions. All the images have been manually labeled considering the following 8 categories: Appetizer, Main Course, Second Course, Single Course, Side Dish, Dessert, Breakfast, Fruit. Images depicting mixed food (e.g., fish with salad), are labeled with multiple labels (e.g., Second Course and Side Dish).

UNICT Food Dataset 889
The UNICT-FD889 dataset is composed by 889 distinct plates of food. This is a first reduced version of UNICT Food Datasets. It has been used to perform some experiments for a proper comparison with previous results.
The figure below shows a sample of the food images which compose the UNICT-FD889 dataset.

Visual Analysis
In order to better understand the retrieval results a visual inspection can be done.
Click for the visual analysis of the three experiments runs: run 1 - run 2 - run 3
In the following different views of results related to the second run are proposed.