You can manage bookmarks using lists, please log in to your user account for this.
Media type:
Doctoral Thesis;
Electronic Thesis;
E-Book
Title:
Learning-based 3D human motion capture and animation synthesis
Contributor:
Habibie, Ikhsanul
[Author]
Published:
Saarländische Universitäts- und Landesbibliothek, 2023
Language:
English
DOI:
https://doi.org/10.22028/D291-40012
Origination:
Footnote:
Diese Datenquelle enthält auch Bestandsnachweise, die nicht zu einem Volltext führen.
Description:
Realistic virtual human avatar is a crucial element in a wide range of applications, from 3D animated movies to emerging AR/VR technologies. However, producing a believable 3D motion for such avatars is widely known to be a challenging task. A traditional 3D human motion generation pipeline consists of several stages, each requiring expensive equipment and skilled human labor to perform, limiting its usage beyond the entertainment industry despite its massive potential benefits. This thesis attempts to explore some alternative solutions to reduce the complexity of the traditional 3D animation pipeline. To this end, it presents several novel ways to perform 3D human motion capture, synthesis, and control. Specifically, it focuses on using learning-based methods to bypass the critical bottlenecks of the classical animation approach. First, a new 3D pose estimation method from in-the-wild monocular images is proposed, eliminating the need for a multi-camera setup in the traditional motion capture system. Second, it explores several data-driven designs to achieve a believable 3D human motion synthesis and control that can potentially reduce the need for manual animation. In particular, the problem of speech-driven 3D gesture synthesis is chosen as the case study due to its uniquely ambiguous nature. The improved motion generation quality is achieved by introducing a novel adversarial objective that rates the difference between real and synthetic data. A novel motion generation strategy is also introduced by combining a classical database search algorithm with a powerful deep learning method, resulting in a greater motion control variation than the purely predictive counterparts. Furthermore, this thesis also contributes a new way of collecting a large-scale 3D motion dataset through the use of learning-based monocular estimations methods. This result demonstrates the promising capability of learning-based monocular approaches and shows the prospect of combining these learning-based modules into an integrated 3D ...