[3Q3-J-13] AI application: analysis of physical behaviors in artifacts

[3Q3-J-13-01] Construction of Dataset for Feature Extraction Performance Evaluation using Aerial Photographs

〇Hiroyuki Ohno¹, Ryo Endo¹, Takayuki Nakano¹, Masako Shinoda¹ (1. Geospatial Information Authority of Japan)

Keywords:Feature Extraction, Semantic Segmentation, Aerial Photograph

Several datasets are known as datasets for feature extraction. However, they are targeted to outside Japan and are not necessarily dataset with high regional diversity. Therefore, it is not suitable as an evaluation dataset for feature extraction in mapping performed as survey in Japan. So, we carried out this research with the aim of enabling the evaluation of feature extraction performance for high regional diversity and aerial photograph actually used in mapping in Japan. As a result, a dataset for evaluation was constructed using aerial photographs of 1328 regions taken in Japan since 1967. Furthermore, we compared the evaluation value of the prediction result using our dataset with that using the existing dataset using pix2pix and U-Net, and concluded that our dataset can perform sufficiently reliable evaluation.