UPDATE  = { 2008-08-28 },
author =      {Martinec, Daniel},
supervisor =  {Pajdla, Tom{\'a}{\v s}},
title =       {Robust Multiview Reconstruction},
school =      {Center for Machine Perception, K13133 FEE
               Czech Technical University},
address =     {Prague, Czech Republic},
year =        {2008},
month =       {June},
day =         {20},
type =        {{PhD Thesis CTU--CMP--2008--01}},
issn =        {1213-2365},
pages =       {134},
figures =     {40},
psurl       = {[Martinec-thesis.pdf]},
project =     {1ET101210406, Dur IG2003-2 062, FP6-IST-027113},
annote = {Reconstructing a 3-dimensional (3D) model of a scene from a
  set of 2D images is a fundamental problem in computer vision with
  many applications. The problem can be decomposed into three
  steps. First, some correspondences between pairs of images are found
  and 3D geometries of the image pairs are estimated. Secondly, the
  two-view geometries are fused into a consistent reconstruction of
  all views. Thirdly, having a complete camera calibration, a
  consistent dense model of the scene surfaces can be reconstructed
  using all images. While the two-view camera calibration is a well
  studied problem, the multiview camera calibration remains a
  challenging task. It is also the most crucial step in the scene
  reconstruction as the quality of the resulting dense 3D model is
  fundamentally limited by precision of the multiview camera
  calibration. This thesis studies mainly the problem of multiview
  camera calibration.  The largest difficulty of the problem is
  sparsity of the data which happens when the images are only sparsely
  captured (so-called wide baseline stereo, WBS). Then, the scene
  contains many occlusions, i.e. many points are seen in a few images
  only. The second difficulty of the problem is handling of incorrect
  correspondences (mismatches), thanks to which also non-existent
  pair-wise geometries can be found. Every such geometry must be
  detected and removed to obtain a correct reconstruction. The main
  contribution of the thesis is a technique for multiview camera
  calibration by gluing partial reconstructions. This technique was
  used for uncalibrated cameras to obtain a projective reconstruction
  as well as for partially calibrated cameras to obtain a metric
  reconstruction. The technique works in practical situations,
  i.e. the perspective camera, many (99.9\%) occlusions in scene and a
  not entirely exact correspondence algorithm. The importance of such
  technique lies in that it offers united and elegant way of
  processing correspondences from WBS and sequences.  The presented
  methods exploit all data known about the scene, namely in the same
  way and at once. The core of the methods is a linear algorithm which
  provides a very good reconstruction already before non-linear
  refinement using bundle adjustment. The developed methods embrace
  projective factorization for points and lines, gluing of projective
  pairwise reconstructions, merging metric panoramas and gluing
  pairwise metric reconstructions. Some of the methods are applicable
  for both affine and perspective camera models. The methods are
  suited for large-scale reconstructions (thousands of images). The
  accuracy, applicability and speed of the methods is demonstrated on
  difficult wide baseline image sets whose metric dense
  reconstructions are shown. The presented techniques were used in a
  complete, robust automatic multiview reconstruction pipeline from
  images to a 3D model.},
keywords =    {computer vision, 3D reconstruction, multiple view reconstruction,
               auto-calibration, epipolar geometry, mismatch identification, 
               dense stereo, omnidirectional camera, line reconstruction},