2

I couldn't find the description of the content in the feature descriptor vector. Do the value make some sense or are they only used to identify similar points in another image? Is the vector something like HuMoments-vector, where each position has a meaning?

Thank you!

P.S.: I'm looking for features to classify images (bio-cells). Currently I'm only using the number of key-points extracted by OpenCV as features. It works quite well, but probably using some values from descriptor vectors could be an improvement?

Valentin H
  • 7,240
  • 12
  • 61
  • 111
  • 1
    depends on descriptor used. Descriptor is typically a kind of projection of keypoint neighborhood to high-dimensional space (for example, 128-dimensional). The method of projection can vary - histograms of gradients, frequency bands, etc – Andrey Kamaev Jan 23 '15 at 20:58
  • @AndreyKamaev: Interesting! Is it documented in OpenCV anywhere? Is there a good link for reading? Does it mean, that the descriptor uses only the location of a given key-point (and probably type). I mean, could I create a descriptor at any point I like? – Valentin H Jan 23 '15 at 23:07

1 Answers1

3

A Keypoint is a detected point (by some algorithm) in an image, that is capable of being detected robustly in different scenarios. But detection isn't enough - the detected keypoints have to be matched in the different images. That is what a descriptor is needed for - it describes that keypoint, again using another set of algorithms. How good that description is will depend on a lot of factors, beginning with the type of image itself.

Descriptors computed for points that are not valid keypoints will be meaningless (and may not even be computable), and most OpenCV algorithms discard such keypoints.

Community
  • 1
  • 1
a-Jays
  • 1,182
  • 9
  • 20