Resumen de Sequential bayesian non-rigid structure from motion
This thesis addresses the problem of recovering simultaneously camera motion and the 3D reconstruction of deformable objects from monocular video. We propose several methods to solve this problem in a sequential fashion, frame-by-frame estimation, as the data arrives. Deformable structures appear constantly in our everyday life, from human non-rigid motion (e.g., a smiling face or performing different expressions) to general objects such as flags, clothes, sails, banners, etc. More speculatively, in the medical field such as a beating heart or a bending abdomen, where the problem is particularly challenging. Our research seeks a physics-based method to perform 3D shape recovery in a wide variety of objects with different types of deformation from inextensibility to extensibility, without having to rely on learning data. In addition, our methods can perform also under realistic real-world assumptions allowing large amounts of missing data and measurement noise, they can run in real time at frame rate and can be used from sparse to dense shapes even for strong deformations.
This dissertation presents our contributions in the field of deformable shape and camera motion recovery from a sequence of monocular images. In more detail, we present a novel algorithm where both motion and deformation are ruled by physical dynamic models. An important advantage of this method is that it does not require prior knowledge over material properties since they can be factorized out. We also present a generic estimation framework, eliminating the need of rigid priors, which is normally necessary when physics-based models are used. Finally, we show how the sequential estimation is possible for dense shapes, combining low-rank shape models with temporal and spatial smoothness priors. One of the main advantages of our models is the ability to include physical priors, if they are available. In contrast, we show how to solve the problem when this knowledge fails.
