Warning
This page is located in archive. Go to the latest version of this course pages.

Seeking of Sparse Correspondences

The WBS matcher (code repository) is a tool for seeking a sparse correspondences. The matcher makes use of descriptors of interest points in the image. These IP descriptors are generated independently for each image, and then the descriptors of particular image pair are matched in order to establish pairwise correspondences. So it is recommended to first pre-compute and store descriptors for all the images, and then use them for pairwise matching.

The WBS matcher uses only local image area similarity for matching, there are no other restriction on correspondences. The resulting correspondences typically contain lot of errors (mismatches), they even need not be one-to-one (i.e., a single IP can be member of more pairs). We call these correspondences tentative. For example see figure 1.

Fig. 1: Tentative pairwise correspondences obtained by WBS (shown 20% of 995). Corresponding points have the same colour (but colour is not unique)

Correspondences found by WBS are pairwise in principle (between two images). However, single image can be member of more pairs, of course. Consider the example situation of correspondences in all pairs between the three images {1,2,3}, i.e., in the pairs {1-2,1-3,2-3}. Since the matching in the pairs 1-2 and 1-3 uses the same interest points in the image 1, we can got correspondences, that links a point in the image 1 to a point in the image 2 and to a point in the image 3 as well. See schematic demonstration in figure 2.

Fig. 2: Transitivity of Correspondences

Note that there are four different cases:

  • a correspondence links only one pair (blue),
  • correspondences link two pairs (violet),
  • correspondences link all the three pairs and they are consistent (green),
  • correspondences link all the three pairs and they are contradictory (red).

figure 3 shows correspondences, that links three pairs and are consistent.

Fig. 3: Triple-wise consistent tentative correspondences (108))

Task 2

  1. Pre-compute (and save) interest point descriptors for all the input images.
  2. Estimate (and save) pairwise correspondences in all pairs between input images (so e.g. for 12 input images there is 66 pairs).
  3. Visualize correspondences for a chosen pair, e.g., as in figure 1.
  4. Choose a triple of images and select and visualize correspondences that links all the images and are consistent, as in figure 3.
courses/tdv/labs/2_sparse_correspondences.txt · Last modified: 2019/09/13 11:38 (external edit)