0

please help me to understand this idea from a paper which titled is "Scene Summarization for Online Image Collections" by Ian Simon Noah Snavely Steven M. Seitz, University of Washington.

Computing the Feature-Image Matrix :
We first transform the set of views into a feature-image incidence matrix. To do so, we use the SIFT keypoint detector to find feature points in all of the images in V. The feature points are represented using the SIFT descriptor. Then, for each pair of images, we perform feature matching on the descriptors to extract a set of candidate matches. We further prune the set of candidates by estimating a fundamental matrix using RANSAC and removing all inconsistent matches After the previous step is complete for all images,

we organize the matches into tracks, where a track is a connected component of features. We remove tracks containing fewer than two features total, or at least two features in the same image. At this point, we consider each track as corresponding to a single 3D point in S. From the set of tracks, it is easy to construct the |S|-by-|V| feature-image incidence matrix.

the part which i confused about is the italic one.
how we organize matches into tracks ?
and how to construct feature-image incidence matrix ?

pls help me. . .

go eng chun
  • 73
  • 2
  • 10

1 Answers1

0

Example for 3 images track.

  1. Detect features

  2. Perform matching (1 - 2, 2 - 3). Now you have correspondences FeatureA_img1 = FeatureB_img2, FeatureC_img2 = FeatureD_img3, FeatureE_img1 = FeatureF_img3.

  3. Check, if FeatureA_img1 == FeatureB_img2 AND FeatureB_img2 == FeatureC_img3, than you have the same feature in 3 images. Save it in the array:

    img1 img2 img3 ... imgn FeatureA FeatureB FeatureC ...

Repeat this for all correspondences. The rows in this table are the tracks you are looking for.

old-ufo
  • 2,799
  • 2
  • 28
  • 40
  • thanks, but for making this tracks, if there is 4 images, should I compare images 1 with 3 others and images 2 with 3 others and so on ? so there would be 4 combination 2 iteration ? – go eng chun Dec 22 '13 at 10:02
  • In tracking usually you track only 1-2, 2-3, 3-4, not 2-4 etc. – old-ufo Dec 22 '13 at 15:40
  • i see, so the source images must be filtered first ? cause if image 1 = image 3 but image1 != image2, it wont work. Is it true? can you give me suggestion, how to group images using sift feature ? first, i get sift descriptor, match it, ransac it. . . and the next after still confused. . . . – go eng chun Dec 23 '13 at 01:32
  • Well, yes. You match images pairwise, then build connection graph between images. May be this paper can help you http://www.robots.ox.ac.uk/~vgg/publications/papers/schaffalitzky02.pdf – old-ufo Dec 23 '13 at 19:17
  • i see, it's help me a lot. As my conclusion, it will be better if i use classification than clustering. But due to my time limit, i'll still use clustering for my thesis. Thanks a lot, old-ufo – go eng chun Dec 31 '13 at 10:55