MrX's profile - activity

OpenCV keypoints have a bunch of attributes of which two are octave and size.

1. Is the descriptor sampling pattern size (from e.g. BRISK or FREAK) defined by those two attributes? If yes, does this mean that if I detect the same physical point in the same octave layer by two different detectors, the descriptor will not be identical?
2. Do descriptors always use this size information or is this optional?

cv::calcOpticalFlowPyrLK takes as the first argument the previous image and as a second argument the next image. Instead, you pass optical flow pyramids. Why do you do that?

I checked the source code of the FREAK descriptor and it seems that it I would have to copy the values that are used there. I don't like this solution and decided to rewrite the framework with which I am working.

It's just that the framework I am working with uses the Eigen library to store keypoint and descriptor data. Furthermore, the detector and extractor class are implemented separately so I pass data from the detector to the extractor with Eigen. Unfortunately, it is quite painful to remove data from Eigen vectors and matrices. That's my problem.

I'm working with a framework that relies on keeping the number of keypoints constant before and after descriptor extraction. So far I applied a detection mask such that keypoints are not detected if they are close (e.g. 30 pixel) to the border of the image. Unfortunately, the extractor still removes a few keypoints. For example, I detect keypoints with the SURF detector on several octaves (!) and then use the FREAK extractor. Everything works fine if the number of octaves of the SURF detector is set to 1.

Can I sort out these keypoints by considering their size and octave? I removed all the keypoints with a size larger than 60 pixels but I've got still the same problem.

I want to use either FREAK or BRISK descriptors that should be scale and rotation invariant. I am quite confused about the following points

1. Do FREAK and BRISK need scale and rotation estimation from the detector? I thought that they can provide rotation invariant descriptions and only need scale information from the detector. Is this correct?
2. Which free detectors can I use for this task if speed is also a constraint?

My thoughts so far: I presume that I could use the BRISK detector. I am also asking myself if this is just the Pyramid version of FAST. If this is true I could probably use other detectors by using the Pyramid modification.

These are more than just one question but I all of them are essential for me to properly use these binary descriptors.

I'm using OpenCV 2.4