Thesis Presentations IMG2HRTF [ak discourse]

 

Liebe Studierende, Mitarbeiter und Freunde der Audiokommunikation,

 

hiermit möchte ich die Einladung zur Präsentation zweier Abschlussarbeiten weiterleiten, die aus einer Kooperation der Fachgebiete Computer Graphics und Audio Communication an der TU hervorgegangen sind.

 

Termin ist Do, der 29. Oktober 2020, 14.00.

 

Zum Inhalt der Arbeiten siehe unten.

 

Mit herzlichem Gruß

Stefan Weinzierl

 

***

Hi everyone,

on Thursday, October 29th Tim Fleckenstein and Oliver Weißbarth will present the results of their theses at 2pm (s.t.). 

The titles of the theses are "Generating a 3D morphable model of pinnae" (Tim) and "Realistic rendering of head regions for learning head-related transfer functions" (Oliver). You're invited to join the 2x20min talks and subsequent Q&As:Time: Oct 29, 2020 02:00 PM Amsterdam, Berlin, Rome, Stockholm, ViennaJoin Zoom Meeting
https://tu-berlin.zoom.us/j/68767615837?pwd=UWRUbE1BZlFYTnViNEN4dXhVMFlIdz09Meeting ID: 687 6761 5837
Passcode: 564749

Abstract (Tim):
This paper aims at providing a 3D Morphable Model (3DMM) of the human head and ear. Although
creating such a 3DMM is an interesting endeavour in and of itself, the aim of the model is to
advance research into the calculation of the Head-Related-Transfer-Function (HRTF). The HRTF
captures how sound is being processed by an individual depending on the shape of his or her head
and ears. Subsequently, this function can then be used to personalize sound and create a three
dimensional sound experience for the listener.Abstract (Oliver):
Human’s ability to localise a sound source (localisation) is based on multiple features that are extracted from an incoming audio signal. One component of human spatial audio is the Head-Related Transfer Function (HRTF). It describes a frequency and direction dependent filter that is caused by the diffraction and reflection of sound at the head, ears and torso. Due to different torso, head and ear shapes this filter is different for each person. To replicate this filter using head-phones it needs to be obtained for each individual. This process is called HRTF Individualisation. Classical approaches use physical measurements or numerical simulation of a 3D mesh of the head and ear. Both approaches require specialised hardware, are time consuming and thus expensive.  Therefore machine-learning approaches have been proposed to predict the individualised HRTF directly from a photo of the ear. Such machine-learning approaches are often limited by the small size of available training data sets. We therefore propose a system that uses data synthesis to create randomised head meshes and use numerical simulation to compute the corresponding HRTF. Photo-realistic rendering is then used to obtain a synthetic photo of the ear. We test and evaluate the synthetic data set using multiple machine-learning models.

Best,

Sebastian

Sebastian Koch

Research Assistant

M.Sc. Computer Engineering

Computer Graphics, EECS

Technische Universität Berlin

Marchstr. 23, 10587 Berlin, Germany 

Tel: +49 (0)30 314-73103

s.koch@tu-berlin.de 

www.cg.tu-berlin.de

mme.png

img2hrtf.png