The pinna geometry descriptor: A topologically consistent three-dimensional ear shape representation for...

In spatial audio research, subject-specific three-dimensional (3D) pinna geometry remains a major bottleneck due to inconsistent mesh topologies, varying sampling densities, and arbitrary vertex indexing across available datasets. To address these limitations, this paper introduces the pinna geometry descriptor (PGD) framework, a methodology that transforms raw 3D ear scans into a topologically consistent, dense point-to-point correspondence space suitable for computational modeling and machine learning applications. The framework employs a semi-automated landmarking strategy and a two-stage non-rigid registration pipeline to map input scans onto a statistically representative template mesh. A manifold-learning spatial reindexing procedure ensures that consecutive vertex indices reflect local physical proximity, optimizing the descriptor for architectures reliant on neighborhood aggregations. The framework is validated on the Viking II dataset, and its transferability is evaluated using the Aachen dataset as an external case under distinct acquisition protocols. The descriptor is used in a state-of-the-art geometry-driven head-related transfer function synthesis model. The results indicate that PGD successfully preserves subject-specific anatomical features while providing the structural uniformity required for deep-learning-based acoustic prediction, establishing its utility for scalable and accurate spatial audio personalization.