Section: New Results

A Bayesian Approach to Multi-view 4D Modeling

This paper considers the problem of automatically recovering temporally consistent animated 3D models of arbitrary shapes in multi-camera setups. An approach is presented that takes as input a sequence of frame-wise reconstructed surfaces and iteratively deforms a reference surface such that it fits the input observations. This approach addresses several issues in this field that include: large frame-to-frame deformations, noise, missing data, outliers and shapes composed of multiple components with arbitrary geometries. The problem is cast as a geometric registration with two major features. First, surface deformations are modeled using mesh decomposition into elements called patches. This strategy ensures robustness by enabling flexible regularization priors through inter-patch rigidity constraints. Second, registration is formulated as a Bayesian estimation that alternates between probabilistic datal-model association and deformation parameter estimation. This accounts for uncertainties in the acquisition process and allows for noise, outliers and missing geometries in the observed meshes. In the case of marker-less 3D human motion capture, this framework can be specialized further with additional articulated motion constraints. Extensive experiments on various 4D datasets show that complex scenes with multiple objects of arbitrary nature can be processed in a robust way. They also demonstrate that the framework can capture human motion and provides visually convincing as well as quantitatively reliable human poses. This work was published as journal article in International Journal on Computer Vision (IJCV) [4] .