Impression Estimation Model of 3D Objects Using Multi-View Convolutional Neural Network

Keisuke Sakashita, Kensuke Tobitani, Koichi Taguchi, Manabu Hashimoto, Iori Tani, Sho Hashimoto, Kenji Katahira, Noriko Nagata

Abstract: The ultimate goal of this study is to provide intuitive design support for 3D objects. As a first attempt, we propose a method for estimating impressions of common 3D objects with various characteristics. Although many studies have been conducted to estimate objects’ aesthetics, not enough research has been conducted to estimate the various impressions of objects necessary for design support. The data set of human impressions of 3D objects is constructed based on psychological methods. To account for the variability in people’s ratings, the distribution of ratings is represented by a histogram. By learning the distribution of impression ratings, with the estimation model, we can realize an impression estimation model with high estimation accuracy. In the accuracy validation experiment, the proposed method’s estimated results (estimated impression distribution) showed a moderate to high positive correlation with the distribution of human impressions. In addition, we confirmed that the proposed method has greater estimation accuracy than previous studies and that it captures the tendency for variation in people’s impression evaluations (the global tendency of impression distribution). Furthermore, visual confirmation of the relationship between the estimation results of the constructed impression estimation model and 3D objects suggests that the proposed method is capable of identifying the main physical features associated with impression words, confirming the proposed method’s validity.