1. Introduction
Reconstructing the 3D shape of deformable objects from monocular images, known as Non-Rigid Structure-from-Motion (NRSfM), has applications in domains ranging from entertainment [23] to medicine [20]. It was introduced in [4] by expressing shapes in terms of a low-rank shape-basis. Many variants of this idea have since been proposed with a view to improve reconstruction stability [5], [2], [32], [11], [21], [14]. Over the last decade, physically-inspired NRSfM models [33], [34], [30], [8], [9], [18], [24], [25] have emerged as an attractive alternative. They exploit local surface properties to draw constraints, can handle large deformations and outperform techniques relying on low-rank priors. Unfortunately, most methods in both categories become prohibitively slow as the number of images increases, due of their non-linear complexity, and cannot handle missing data. This makes them impractical for real-world scenarios.