如何在opencv中制作自己的特征检测方法？

Question

Let's take a look on this basic tutorial named Features2D + Homography to find a known object. It uses SurfFeatureDetector to detect features:

  SurfFeatureDetector detector( minHessian );
  std::vector<KeyPoint> keypoints_object, keypoints_scene;
  detector.detect( img_object, keypoints_object );
  detector.detect( img_scene, keypoints_scene );

Then it uses SurfDescriptorExtractor to calculate descriptors (feature vectors) using detected features.

My questions are:

1. if I want to create my own feature detector (for example with Trajkovic or Harris algorithms) which Descriptor Extractor shall I use?
2. are the features, that were found in SurfFeatureDetector, just the common points or the areas of points?

**** addition * ***

1) In this example the Surf algorithm of feature detection was used. I have made my own algorithm (Trajkovic) and it works great - all the corners (image features) are found. Then I try to use SurfDescriptorExtractor as it was used in example. The problem is that SurfDescriptorExtractor don't want to use my founded points in correct way (resulted picture appears with wrong connections, that means, that extractor didn't calculate the vectors correctly).

2) I need to make it exactly using opencv, that's the point;

3) The "feature detector" is an algorithm, that tries to find keypoints (features or corners) on an image, "descriptor extractor" - is an algorithm, that calculates the feature vectors for the best understanding the keypoint position and direction;

4) In conclusion, in example all keypoints are connected on both images (as you can see on the last picture of tutorial) and then highlighted with rectangle. But when I use Trajkovic algorithm, they connected in wrong way, that's why there is no highlighted rectangle.

Answer 1

While we can't pinpoint the problem without looking at your implementation, and possibly even doing some investigating, I can point you in a direction of a solution: the OpenCV source code, you can compare against their implementation.

Have a look at the detectAndCompute() function:

• surf.hpp .
• surf.cpp

Harris corner detection works a bit differently and so is its API:

• corner.cpp