3D Non-Rigid Reconstruction with Prior Shape Constraints

Abstract

3D non-rigid shape recovery from a single uncalibrated camera is a challenging, under-constrained problem in computer vision. Although tremendous progress has been achieved towards solving the problem, two main limitations still exist in most previous solutions. First, current methods focus on non-incremental solutions, that is, the algorithms require collection of all the measurement data before the reconstruction takes place. This methodology is inherently unsuitable for applications requiring real-time solutions. At the same time, most of the existing approaches assume that 3D shapes can be accurately modelled in a linear subspace. These methods are simple and have been proven effective for reconstructions of objects with relatively small deformations, but have considerable limitations when the deformations are large or complex. The non-linear deformations are often observed in highly flexible objects for which the use of the linear model is impractical.
Note that specific types of shape variation might be governed by only a small number of parameters and therefore can be well-represented in a low dimensional manifold. The methods proposed in this thesis aim to estimate the non-rigid shapes and the corresponding camera trajectories, based on both the observations and the prior learned manifold.
Firstly, an incremental approach is proposed for estimating the deformable objects. An important advantage of this method is the ability to reconstruct the 3D shape from a newly observed image and update the parameters in 3D shape space. However, this recursive method assumes the deformable shapes only have small variations from a mean shape, thus is still not feasible for objects subject to large scale deformations. To address this problem, a series of approaches are proposed, all based on non-linear manifold learning techniques. Such manifold is used as a shape prior, with the reconstructed shapes constrained to lie within the manifold. Those non-linear manifold based approaches significantly improve the quality of reconstructed results and are well-adapted to different types of shapes undergoing significant and complex deformations.
Throughout the thesis, methods are validated quantitatively on 2D points sequences projected from the 3D motion capture data for a ground truth comparison, and are qualitatively demonstrated on real example of 2D video sequences. Comparisons are made for the proposed methods against several state-of-the-art techniques, with results shown for a variety of challenging deformable objects. Extensive experiments also demonstrate the robustness of the proposed algorithms with respect to measurement noise and missing data.