A Data-Driven Parametric Framework for Size-Adaptive Shoe Insole Outline Generation

With the rapid expansion of AR/VR-based digital platforms, there is an increasing demand for the automated generation of size-varied 3D shoe assets. However, conventional CAD-based linear scaling and PCA-based global statistical shape models are limited in their ability to capture the non-uniform and locally non-linear deformations observed in insole contours. This study proposes a size-adaptive insole contour generation framework that integrates image-based contour extraction, B-spline parametric representation, and type-specific SVR-based local displacement regression. By decomposing control-point displacements into tangent–normal components, the proposed method directly models non-linear curvature variations associated with size progression without relying on dimensionality reduction. Quantitative evaluations under an eight-fold leave-one-insole-out (LOIO) protocol show that the proposed method achieves a mean Hausdorff distance of 4.73 mm, a Chamfer distance of 1.76 mm, and an IoU of 0.929. It significantly outperforms the no-clustering ablation in both the Hausdorff distance (6.13 mm; p=0.032) and IoU (p=0.033), as well as the PCA-based kernel ridge regression baseline across all three metrics (p<0.05). No statistically significant differences were observed between the proposed method and the ratio scaling, Gaussian process, or thin plate spline baselines (p>0.05). PCA-Linear showed a small numerical advantage in the Hausdorff distance (4.16 mm), but the difference was not statistically significant (p=0.187). A sensitivity analysis further reveals the existence of a practical control-point density region that balances geometric fidelity and model complexity. This work provides a data-driven parametric approach for standard size-based digital grading automation and establishes a technical foundation for future extension to full 3D shoe mesh generation.