University of Thessaly (UTH)
AI-based synthesis for complex human poses, motions, and interactions
Abstract
dc:descriptionMotion Capture (MoCap) is a technology with broad applications, yet its democratization remains an open research topic. Even high-end optical systems require laborious manual cleanup, while recent advances in lower-cost, markerless MoCap are hampered by poor data quality, making them almost useless for downstream applications. This Thesis addresses these challenges, taking a shift from existing solutions and leveraging AI synthesis, enabling the accurate capture of complex poses, motion, and interactions. First, we tackle the challenge of complex human poses by using representation learning to synthesize new training data, balancing existing MoCap datasets to train more effective AI models. We demonstrate the efficacy of this approach on the task of automatic marker labeling, a critical step in optical MoCap workflows. However, these models can introduce complex noise patterns, which, combined with low-cost sensors often used in real-world settings, lead to significant uncertainty in measurements. Existing optimization approaches often assume clean data or simple noise models, making them ill-suited for these scenarios. We, therefore, propose a novel optimization framework that models the uncertainty of the constraints themselves, learning it alongside the measurements. While effective for optical MoCap, this method is insufficient for the far more challenging case of markerless data. Markerless MoCap suffers from severe artifacts, including jittery joint estimates, swapped body parts, and completely missing data, which are intractable for most solvers. Furthermore, most approaches fail to leverage the temporal coherence present in motion data. Hence, we propose a framework for robustly capturing motions, which leverages a learned manifold with specific geometric properties to represent the space of valid human poses. This enables us to introduce novel synthesis techniques that inherently leverage temporal coherence to enable efficient motion solving while effectively alleviating severe artifacts. In summary, this Thesis presents a suite of innovative techniques that utilize representation learning to synthesize new samples for balancing training data, modelcomplex noise patterns, and robustly solve challenging motion capture scenarios, thereby taking a step towards making high-quality MoCap widely accessible.
Degree
thesis:*- Grantor dc:publisher
- University of Thessaly (UTH)
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Almpanis, Georgios
- Αλμπάνης, Γεώργιος
Subjects
dc:subject × 12Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/61158