Author: Maxime Lhuillier
Year: 2002
- quasi-dense matching algorithm: starts with sparse sets of matches and propagates the neighbouring pixels
- IBMR (image base modeling and rendering) = 3D reconstruciton + rendering : the aim is to synthetize new views
- Can be done with implicit 3D reconstruction with essential matrix or trifocal tensor
- "dense stereo algorithms are not sufficiently robust except for precalibrated cameras in presettled environment "
- "sparse matching has been shown to be efficient due to its highly discriminant nature of the points of interest "
- extract GFTT and match them with ZNCC using cross consistency check (ZNCC > 0.8 w/ 11 * 11 windows)
- Then propagate each match in their spatial neighborhood starting with the matches with the best ZNCC score (best first strategy)
- Generalize the 1D disparity gradient limit to 2D i.e. matched pixels must have matches in their surroundings with a similar 2D distance (use 5x5 neighbourhood for this check)
- perform a confidence measure to get rid of uniform areas
- Each time a new match is declared, it is added to the seed (avalanche behavior)
- The complexity of the algorithm is output dependant (i.e. it depends on the number of matched pixels)
- Avalanche behavior: needs a few seed matches
- Works well on low textured images: propagates along the lines