Contributions to Spatial and Temporal 3-D Reconstruction from Multiple Cameras

This thesis addresses the necessary steps required to build a framework for spatial and temporal 3-D reconstruction using multiple camera environments: camera calibration and sparse 3-D reconstruction, dense 3-D reconstruction, sparse and dense mesh matching. Firstly, a probabilistic formulation is developed in conjunction with any affine factorization algorithm (3-D point based reconstruction method based on matrix factorization), able to recover both the extrinsic parameters of multiple cameras and 3-D coordinates of control points, given their projected 2-D point correspondences and the intrinsic camera parameters. The proposed framework is robust to outliers and compares favourably with bundle adjustment, a standard non-linear minimization technique, which requires an initial solution not very far from the optimum. Secondly, a provably correct mesh-based surface evolution approach is proposed. It is able to handle topological changes and self-intersections without imposing any mesh sampling constraints. The exact mesh geometry is preserved throughout, except for the self-intersection areas. Sample applications, including mesh morphing and 3-D reconstruction using variational methods, are presented. Thirdly, a scene-aware camera clustering method is developed, able to break large-scale reconstruction tasks in smaller independent partial reconstructions that are memory tractable. Lastly, a new 3 dimensional descriptor is proposed, defined on uniformly sampled triangular meshes. It is invariant to rotation, translation, scale, being able to capture local geometric and photometric properties. It is particularly useful in the multi-camera environments, where the reconstructed meshes benefit from colour information. Nevertheless, the descriptor is defined generically for any feature available throughout the manifold, colour and curvatures being just some examples. Results in both rigid and non rigid matching tasks are presented. Additionally, the descriptor is integrated within a mesh tracking framework, providing dense matches.

@PhdThesis\{Zah08,
  author       = "Zaharescu, Andrei",
  title        = "Contributions to Spatial and Temporal 3-D Reconstruction from Multiple Cameras",
  school       = "Institut National Polytechnique de Grenoble",
  month        = "November",
  year         = "2008",
  address      = "Grenoble, France",
  url          = "http://perception.inrialpes.fr/Publications/2008/Zah08"
}