Learning to Rank Food Images

D. Allegra, D. Erba, G. M. Farinella, G. Grazioso, P. D. Maci, F. Stanco, V. Tomaselli

International Conference on Image Analysis and Processing, Trento, Italy, September 2019

Download paper

Giovanni Maria Farinella

We address the problem of volume estimation through Learning to Rank algorithms.

The proposed dataset consists of 99 RGB images belonging to 11 different classes. Each image is associated to one portion size among three possible portion sizes: small, medium and large. For each class we collected 9 images corresponding to 3 images for each portion size. Moreover, in order to introduce variability, during acquisition we have used plates with two different diameters: 18cm and 22.8cm. All the acquisition have been performed by a standard RGB camera fixed in a support and a centered top view with respect to the plate. The proposed dataset includes multiple portion sizes for each dish to properly test LTR methods. The proposed dataset is publicly available to promote new task of ranking food images and to make repeatable our experiments.