畳み込みニューラルネットワークを用いた自動車の三次元モデルにおける各車型の特徴抽出と視覚化

抄録

The cars are classified by cars' body types. However the characteristics are basically similar at first sight, so it is difficult to distinguish the differences among those cars' body types. Therefore, in this study, we considered that cars' characteristics could be analyzed by using deep learning and image recognition technology, developed a system to visualize the judgment and characteristic parts of cars' body types. Specifically, we made renderings of the CG model of 30 cars by setting 360 viewpoints in 1 degree increments around each car. Deep learning was performed using these 2D images as teacher signals. The car body type recognition probability of each angle is graphed, and the characteristic parts of each car body type are visualized. As a result, we clarified the visual angles and the pars contributing the judgment of cars' body types.