IS = { zkontrolovano 15 Jan 2008 },
  UPDATE  = { 2008-01-14 },
author =      {Kostliv{\'a}, Jana},
supervisor =  {{\v S}{\'a}ra, Radim},
title =       {Stratified Dense Matching for Stereopsis in Complex Scenes},
school =      {Center for Machine Perception, K13133 FEE
               Czech Technical University},
address =     {Prague, Czech Republic},
year =        {2008},
month =       {January},
day =         {},
type =        {{PhD Thesis CTU--CMP--2007--26}},
issn =        {1213-2365},
pages =       {129},
figures =     {45},
authorship =  {100},
psurl       = {[Kostliva-TR-2007-26.pdf]},
project =     {1ET101210406, IST-2001-32184, MSM 212300013, GACR 102/01/1371, 
               CTU 0209113,  CTU 0306413, Dur IG2003-2 062, Aktion 34p24,  
               CONEX GZ 45.535 },
annote = {One of the important tasks of computer vision is automatic
  3D scene reconstruction from a given set of images. The fundamental
  part of this task lies in establishing dense correspondences between
  the images. The matching should be error free (in terms of
  mismatches and false-positives) to attain the reconstruction
  successfully, since these kinds of errors may spoil the
  reconstruction process completely.  These requirements mean that not
  the whole of the images should be matched. On the contrary, only
  those parts of the input images, which are unambiguous, should be
  interpreted.  To achieve that, the matching method must be robust,
  i.e.  to be able to reject unreliable data. The quality of the
  results of these methods (typically) depends on the quality of
  matching features, since the matching problem is solved directly
  over them. We present an approach for matching feature
  modelling. The goal is to achieve discriminability and invariance to
  image distortion due to the projection of 3D shape to two different
  vantage points. Such features must therefore be defined in 3D space,
  disparity space, to be able to adapt to original scene surface and
  thus to cover the same part of the scene. This is unlike the
  standard features defined in input images. Our matching features
  adapt to disparity components, high-similarity matching hypotheses
  in disparity space. However, such features cannot be found without
  having matching hypotheses first, thus we split the matching process
  into two semi-independent parts, where in the first one, the
  reliable matching features are modelled over a subset of the
  discrete disparity space, while in the second one, the final
  matching is established. In a ground-truth based evaluation, we have
  shown that proper matching feature mod- elling is able to improve
  matching accuracy 2.3x, matching density 3x, and occlusion boundary
  detection 2x over the standard fixed-size rectangular features
  defined in input images. Experiments on real outdoor complex scenes
  demonstrated the suitability of our approach for solving the
  matching in this kind of scenes. Finally, we present the application
  of our method to 3D scene reconstruction with satisfactory
keywords =    {stereopsis, matching},