[4F3-GS-10] AI application

[4F3-GS-10-04] Estimation Model for Subjective Impression Evaluation of Building’s Exterior Appearance Using Big Data of Street Omnidirectional Image

〇Yusuke Imadegawa¹, Takuya Oki¹, Yoshiki Ogawa², Chenbo Zhao² (1. Tokyo Institute of Technology, 2. The University of Tokyo)

Keywords:Impression evaluation, Image big data, Building exterior appearance, Deep learning, Questionnaire survey

The building's exterior appearance is a crucial element influencing the building's overall impression and the cityscape. However, there have been few quantitative and versatile methods for evaluating the impression of the building exterior, and designers have often relied on their personal experience and intuition. This paper proposed the estimation model for subjective impression evaluation of the building's exterior appearance. First, we extracted 1,000 exterior images using computer vision methods from big data of omnidirectional street images in Ota Ward, Tokyo. Next, we conducted a large-scale web questionnaire about impression evaluation using these images. We used the questionnaire results for training the deep learning model, which enabled us to estimate the impression score of each image based on 22 items. The accuracy of the trained model achieved 78.6% on average. In addition, we demonstrated several primary analyses of the relationship between the building exterior and the impression evaluation score estimated by the trained model.