TopoCap: Learning Topology-Agnostic Motion Priors for Monocular Video-to-Animation
Researchers have developed TopoCap, a novel framework for generating animations from monocular video that can adapt to any skeletal structure. This system learns a universal motion manifold, disentangling motion dynamics from specific topologies. It utilizes a Graph CVAE and conditional flow matching to predict topology-agnostic motion codes from visual input. The framework was trained on Mobjaverse, a large-scale dataset featuring over 5,000 skeletal topologies, enabling zero-shot retargeting for diverse 3D characters. AI
IMPACT Enables animation of arbitrary 3D characters from video, potentially streamlining content creation for games and VFX.