Facial Motion Capture with Sparse 2D-to-3D Active Appearance Models
Author
Plenge, Esben
Term
4. term
Publication year
2008
Pages
96
Abstract
This thesis presents a framework for markerless 3D facial motion capture, meaning it tracks expressions without physical markers on the skin. It focuses on whether emotions can still be conveyed when the 3D data is sparse, meaning it contains only a limited set of key points rather than a dense, detailed model. To enable this, a 2D-to-3D Active Appearance Models (AAM) method is developed to recover sparse 3D geometry from 2D images. The framework is modular: a Matlab tool collects 3D facial geometry to train a statistical model; a C++ method synthesizes 3D geometry from 2D images; and a C++ 3D animation front end visualizes the captured face data. The work uses the AAM-API, an open-source C++ implementation of Active Appearance Models. The thesis describes the underlying theory and the implementation in detail. The framework's ability to convey emotion is tested, and the results show that emotions embedded in the captured and animated facial expressions can be interpreted correctly.
Dette speciale præsenterer en ramme for markørløs 3D-ansigtsregistrering af bevægelse, altså sporing af ansigtsudtryk uden fysiske markører på huden. Vi undersøger især, om følelser kan formidles, når 3D-data kun består af sparsom geometri (et begrænset sæt nøglepunkter frem for en tæt, detaljeret model). Til dette udvikles metoden 2D-til-3D Active Appearance Models (AAM) til at hente sparsom 3D-geometri ud fra 2D-billeder. Rammeværket er modulært: et Matlab-værktøj indsamler 3D-ansigtsgeometri til at træne en statistisk model; en C++-baseret metode syntetiserer 3D-geometri fra 2D-billeder; og en C++-baseret 3D-animationsfrontend visualiserer de indfangede ansigtsdata. Arbejdet benytter AAM-API, en open source C++-implementering af Active Appearance Models. Specialet giver en grundig gennemgang af teorien bag og diskuterer implementeringen i detaljer. Rammeværkets evne til at formidle følelser testes, og resultaterne viser, at de følelser, der ligger i de indfangede og animerede ansigtsudtryk, kan fortolkes korrekt.
[This apstract has been rewritten with the help of AI based on the project's original abstract]