📝 SummarXiv

Abstract

Traditional Head-Related Transfer Functions (HRTFs) acquisition methods rely on specialised equipment and acoustic expertise, posing accessibility challenges. Alternatively, high-resolution 3D modelling offers a pathway to numerically synthesise HRTFs using Boundary Elements Methods and others. However, the high cost and limited availability of advanced 3D scanners restrict their applicability. Photogrammetry has been proposed as a solution for generating 3D head meshes, though its resolution limitations restrict its application for HRTF synthesis. To address these limitations, this study investigates the feasibility of using Graph Neural Networks (GNN) using neural subdivision techniques for upsampling low-resolution Photogrammetry-Reconstructed (PR) meshes into high-resolution meshes, which can then be employed to synthesise individual HRTFs. Photogrammetry data from the SONICOM dataset are processed using Apple Photogrammetry API to reconstruct low-resolution head meshes. The dataset of paired low- and high-resolution meshes is then used to train a GNN to upscale low-resolution inputs to high-resolution outputs, using a Hausdorff Distance-based loss function. The GNN's performance on unseen photogrammetry data is validated geometrically and through synthesised HRTFs generated via Mesh2HRTF. Synthesised HRTFs are evaluated against those computed from high-resolution 3D scans, to acoustically measured HRTFs, and to the KEMAR HRTF using perceptually-relevant numerical analyses as well as behavioural experiments, including localisation and Spatial Release from Masking (SRM) tasks.