On March 14, Thomas Besnier from Lille University, France, working on developing algorithms and new methods for 3D/4D generative models of non-rigid shapes, visited Copenhagen and presented a talk at the weekly Stochastic Morphometry seminar.
The talk was dedicated to computing meaningful deformations of surface meshes, which is crucial for applications ranging from animation to medical imaging. Traditional optimization techniques offer a theoretical framework useful to compute statistics for example. However, they often lack computational efficiency and require extensive parameter tuning, hindering their practical use. Recent deep learning approaches address this issue, but they are limited to registered meshes that have fixed resolution and vertex-wise correspondence. This drastically reduces their robustness and utility when registered data is scarce. In his talk, Thomas delved into these challenges and explained how to empower neural deformation models to operate effectively on unregistered meshes. This new paradigm broadens the range of applicability of deep deformation models with various practical applications. Additionally, he highlighted key open research directions and challenges to build better and more generalizable deformation models.