Abstract

Existing approaches to non-rigid 3D reconstruction either are specifically designed for feature point correspondences, or require a good shape initialization to exploit more complex image likelihoods. In this paper, we formulate reconstruction as inference in a graphical model, where the variables encode the rotations and translations of the facets of a surface mesh. This lets us exploit complex likelihoods even in the absence of a good initialization. In contrast to existing approaches that set the weights of the likelihood terms manually, our formulation allows us to learn them from as few as a single training example. To improve efficiency, we combine our structured prediction formalism with a gradient-based scheme. Our experiments show that our approach yields tremendous improvement over state-of-the-art gradient-based methods.