Neural Deformation Graphs for Globally-Consistent Non-Rigid Reconstruction

Figure 1 : Neural Deformation Graphs: given range input data, represented as a signed distance field, our method predicts globally-consistent deformation graph that is used to reconstruct the non-rigidly deforming surface of an object. The surface of the object is represented as a set of implicit functions centered around the deformation graph nodes. Our global optimization provides consistent surface and deformation prediction, enabling robust tracking of an observed input sequence and even multiple disjoint captures of the same object (as we do not assume sequential input data).