University of Crete (UOC)
Hand texture priors for inverse rendering of hand images in the wild
Abstract
dc:descriptionThis PhD dissertation addresses the challenge of reconstructing accurate 3D hand geometry and texture from unconstrained RGB images, a fundamentally ill-posed inverse rendering problem. Human hands are uniquely difficult to capture due to their highly articulated nature, frequent self-occlusion, and complex appearance, making monocular 3D hand reconstruction extremely challenging. Overcoming these limitations is crucial, as robust vision-based hand capture would enable a wide range of applications, including immersive AR/VR interactions, markerless motion capture, fine-grained gesture recognition (e.g., sign language interpretation) and telemedicine. To tackle these challenges, the dissertation proposes a staged framework that progressively relaxes input constraints from multi-view to monocular setups while incorporating increasingly strong learned priors. First, a Static Multiview Refinement (SMVR) method uses multiple calibrated views and differentiable rendering to refine hand pose, shape, and appearance so that rendered model outputs are photometrically consistent with all observed images. Second, Dynamic Multiview Refinement (DMVR) extends this approach to multiview video sequences, enforcing temporal coherence across frames and disentangling texture from illumination for smooth, consistent reconstructions over time. Finally, a Monocular Texture-Guided Reconstruction framework addresses the extreme ambiguity of single-image hand capture by introducing a learned texture prior trained on in-the-wild data.In a fully differentiable pipeline, the monocular method jointly estimates hand pose, shape, and texture from a single RGB image, achieving high geometric fidelity that surpasses current baselines. Overall, by fusing photometric consistency, temporal dynamics, and learned texture priors, this work enables robust, high fidelity 3D hand reconstruction in-the-wild, bridging the gap between controlled setups and real-world applications.
Degree
thesis:*- Grantor dc:publisher
- University of Crete (UOC)
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Karvounas, Georgios
- Καρβούνας, Γεώργιος
Subjects
dc:subject × 16- Γεωμετρία πολλαπλών όψεων
- Τρισδιάστατη ανακατασκευή χεριού
- Διαφορίσιμη απόδοση
- Εκ των προτέρων γνώση υφής χεριού
- Static refinement
- Temporal refinement
- Multiview geometry
- 3D hand reconstruction
- Differentiable rendering
- Hand texture priors
- Επιστήμη Ηλεκτρολόγου Μηχανικού, Ηλεκτρονικού Μηχανικού, Μηχανικού Η/Υ
- Επιστήμες Μηχανικού και Τεχνολογία
- Ρομποτική
- Electrical Engineering, Electronic Engineering, Information Engineering
- Engineering and Technology
- Robotics
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/62322