Learning Generalizable Representations for Digital Humans and Animals

The modeling of digital humans and animals poses a fundamental challenge in computer vision and computer graphics: how can we learn representations that generalize across heterogeneous skeletal topologies, diverse body shapes, and even distinct species? Most existing approaches remain constrained to fixed skeleton definitions, human-centric datasets, or specific capture setups. Recent advances in motion capture, generative modeling, and volumetric reconstruction have significantly improved realism but this fragmentation still limits scalability and prevents the development of unified frameworks for articulated digital beings.
This seminar discusses recent advances toward topology- and species-agnostic representations for articulated characters. Attention is given to self-supervised strategies for skeleton unification, topology-aware motion modeling, cross-species motion synthesis, and shape-consistent retargeting. The role of high-fidelity geometry and volumetric capture is also examined, highlighting how representation learning must extend beyond motion to include appearance and capture system design.
Such generalizable representations have been proven to improve robustness across heterogeneous skeletons, reduce the dependency on dataset-specific assumptions, and enable scalable modeling of articulated humans and animals, ultimately paving the way toward unified digital character modeling independent of topology, shape, or species.