2026 年 14 巻 4 号 p. 192-200
In Spatial Augmented Reality (SAR), conventional projection mapping often treats target objects as mere screens, potentially diminishing their inherent value. To address this, we propose “Material Transformation.” This approach changes the material appearance of an object while preserving its original physical structures, such as boundaries and joints.
To realize this, we generate textures using NoiseNCA. We evaluate two methods to preserve structural features: integrating edge images into texture generation and using region segmentation. Furthermore, reproducing texture gradients is essential for perceiving the 3D scene structure. We examine three methods for this: using illuminance gradients, area ratios, and a depth estimation model.
Quantitative evaluations show that the region segmentation method is superior for preserving structural features. For reproducing texture gradients, each method has specific suitable and unsuitable scenes. We clarify their effectiveness and limitations to realize Material Transformation effectively.
