Digital Humans: Improving Appearance Rendering, Advancing Motion Retargeting

The concept of a Digital Human refers to the creation, simulation, and interaction with lifelike virtual human representations within digital environments. Digital Humans play a pivotal role across diverse domains, including entertainment, virtual reality, healthcare, and education, enabling applications such as immersive training simulations, interactive virtual assistants, and advanced medical modeling. The definition of a Digital Human is grounded in the analysis of two core aspects: motion modeling and appearance rendering. Motion modeling aims at understanding and replicating human kinematics. The standard representation of humans' skeleton consists of multiple joint-based rigid kinematic model. Such representation differs across the available models, limiting interoperability. This leads to tedious motion retargeting between standards, requiring the manual re-association of the joints. On the other hand, appearance rendering aims to achieve highly realistic visual representations of humans by capturing fine-grained details of geometry and texture. Creating high-quality representation of human appearance requires leveraging multiple data modalities, such as RGB images providing appearance information, or depth maps and point cloud providing 3D geometry information. This thesis addresses these challenges through two primary contributions. First, a solution is proposed to enhance skeleton retargeting by enabling the automatic animation of characters regardless of the specific skeleton representation; additionally, a novel collision optimization process is introduced, ensuring that the retargeted motion is both realistic and free of interpenetrations. Second, the scarcity of comprehensive 4D human datasets is tackled by presenting a large-scale dataset of high-resolution textured meshes, enriched with extensive ground truth annotations. This dataset bridges a critical gap in the literature, offering the first large scale high-quality virtual human dataset collected using a specialized volumetric capture system. As a result, the meshes obtained are significantly more detailed compared to those produced by state-of-the-art methods. Finally, the versatility of the proposed methods is showcased, through applications across diverse domains, demonstrating their potential.