Localized Shape Modelling with Global Coherence: An Inverse Spectral Approach

Many natural shapes have most of their characterizing features concentrated over a few regions in space. For example, humans and animals distinctive head shapes, while inorganic objects like chairs airplanes are made well-localized functional parts with specific geometric features. Often, these strongly correlated – modification facial traits quadruped should induce changes to the body structure. However, shape modelling applications, types edits among hardest ones; they require high precision, but also global awareness entire shape. Even deep learning era, obtaining manipulable representations that satisfy such requirements is an open problem posing significant constraints. In this work, we address by defining data-driven model upon family linear operators (variants mesh Laplacian), whose spectra capture local properties at hand. Modifications translated semantically valid deformations corresponding surface. By explicitly decoupling from surface features, our pipeline allows perform simultaneously maintaining stylistic coherence. We empirically demonstrate how learning-based generalizes not seen training time, systematically analyze different choices diverse categories.