0

I am working on my implementation of the ORB descriptor. I went trough the paper but I am finding hard to understand how to select the training set that is used on their learning method to select a good subset of binary tests.

If I have a single image with very few keypoints shall I use as training set all the patches of the keypoints at hand or only the patches of the keypoint I want to describe?

Andres Felipe
  • 625
  • 7
  • 24

1 Answers1

1

In a more careful reading of the paper I found out that there are two preliminary steps before the algorithm starts:

  1. Define the training set of keypoints from one or several images.
  2. Enumerate all possible tests (pairs of test locations of size 5x5) in a window of size 31x31 (that is they are the same for all training patches).

If there is only one keypoint the algorithm makes not much sense since running all enumerated tests on the keypoint gives 1-length binary string per test and ordering them by its mean doesn't achieve the desired discriminative ordering before the greedy search to chose the less correlated tests.

Andres Felipe
  • 625
  • 7
  • 24