Synthesis of realistic food dataset using generative adversarial network based on RGB and depth images

Al Aama Obada, Tamukoh Hakaru

doi:10.5954/icarob.2021.os19-4

Synthesis of realistic food dataset using generative adversarial network based on RGB and depth images

DOI オープンアクセス

Al Aama Obada

Department of Life Science and Systems Engineering, Graduate School of Life Science and Systems Engineering, Kyushu Institute of Technology
Tamukoh Hakaru

Department of Human Intelligence System, Graduate School of Life Science and Systems Engineering, Kyushu Institute of Technology

説明

Constructing a large food dataset is time and effort consuming due to the need to cover the feature variations of food items. Hence, a huge data is needed for training neural networks. This paper aims to advocate the Cycle-GAN to build up large food dataset based on large number of simulated images and relatively few real captured images thus obtaining more realistic images effortlessly compared with traditional capturing. Real RGB and depth images of variant food samples allocated over turntable were captured in three different angles using real-sense depth camera with different backgrounds. Furthermore, for simulated images, the Autodesk 3D_Maya software was employed using the same parameters of captured real images. Results showed that generally, realistic style transfer on simulated food objects was obtained as a result of employing Cycle-GAN. GAN proved to be an efficient tool that could minimize imaging efforts resulting in realistic images.