Prediction of parameters of a pinna model from synthetic geometries using a vision transformera)

The acquisition of the human pinna geometry requires elaborate equipment for accurate results. Even then, the results are often corrupted by measurement artifacts. We introduce Mesh2PPM, a framework facilitating the generation of a personalized and artifact-free pinna mesh. Mesh2PPM predicts the parameters of a parametric pinna model based on cubic Bézier curves (BezierPPM) from a pinna mesh via a vision transformer. We evaluated Mesh2PPM with multi-view renderings of synthetic pinna geometries, providing additional depth information, varying the grids of camera views, and jittering the camera views. While added depth information had no practically relevant effect, a grid with 3×3 camera views facilitated the lowest overall prediction errors and best counteracted the detrimental effects of jitter. For this grid, with and without jitter, the median Pompeiu-Hausdorff distances were 1.98 mm and 1.34 mm, respectively, and the root mean square distances were 0.92 mm and 0.52 mm. A refined analysis targeting the perceptually most important pinna regions for sound localization showed that multi-view information particularly improved the prediction of BezierPPM parameters describing the cavum-conchae region. The accuracy achieved indicates the suitability of Mesh2PPM to retrieve BezierPPM parameters from pinna meshes.