University of York
Towards an intelligent agent for the human face: Improving the accuracy, controllability, and explainability of 3D face reconstruction
Abstract
dc:description.abstractThe human face is a critical cue for human interaction, playing an essential role in recognition, communication, and even medical diagnosis. However, 3D face reconstruction from a 2D image is an ill-posed problem, with existing approaches struggling due to natural variation in facial appearance and limited 3D data. This thesis addresses these limitations by proposing the Intelligent Face Agent (IFA), which envisions a new form of computational interaction with human faces. The IFA accepts multi-modal inputs and offers intuitive, text-driven manipulation and explanation of 3D facial reconstructions. We use this concept to motivate the design and implementation of complementary components in 3D face generation and analysis. First, we introduce the SynthFace Generator, a novel approach for fast, large- scale 2D-3D paired dataset generation. This method eliminates the need for manual asset creation, producing photorealistic face images with paired 3D shapes. We use this method to create SynthFace, the largest paired 2D-3D dataset for human face shape. Secondly, we present Text2Face, the first method to enable the direct and complete initialisation of 3D face models from textual descriptions. This enhances the controllability of 3D face reconstruction, offering the opportunity to improve practical applications such as avatar creation. Finally, we introduce new baselines for evaluating 3D face reconstruction methods. We propose OptiFaces, a novel baseline that assesses the performance achievable by accurately classifying a set of well-distributed reference faces, providing a more meaningful interpretation of reconstruction error. Additionally, we introduce "N Heads Are Better Than One", a new approach for evaluating combinations of existing 3D face reconstruction methods, resulting in a range of robust new baselines for 3D face reconstruction. The research presented improves the accuracy, controllability, and explainability of 3D face reconstruction, paving the way for broader adoption and application in fields such as healthcare, security, and the creative industries.
Degree
thesis:*- Name dc:type.qualificationname
- Ph.D
- Level dc:type.qualificationlevel
- doctoral
- Grantor dc:publisher.institution
- University of York
- Year dc:date.issued
- 2024
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Rowan, William James
- Advisors dc:contributor.advisor
-
- Huber, Patrik
- Pears, Nick
- Keeling, Andrew
- ORCID person:identifier.orcid
Rights
- Language dc:language.dcterms_iso639_2
- eng
Identifiers
dc:identifier.*- Dc Identifier Grantnumber
- EP/T518025/1