Hi,
I'm currently a phd student. I'm working matching with descriptors 2D and I'm using the evaluation criterion proposed by Mikolajczyk and Schmid, "features2d/evaluation.cpp" file from OpenCV. I'm working with a rgbd dataset so I haven't a perfect homography, the scene isn't coplanar, however I have the trajectory information so I can use this information to perform the transformation of one point from one image to another.
The problem is that I'm evaluating differents descriptors SIFT, SURF and ORB and I have observed that SIFT produces Keypoints with lower size than the other detectors, it means lower ellipse regions. Consequently the evaluation criterion, specifically the IoU step, is more difficult to achieve with this detector than the other whose elliptical regions are bigger and finally recall-precision values are worse.
I don't know if I'm doing something wrong, or there is some mechanism that values the scale level in which the keypoint was detected to enlarge the keypoint size or something?
Maybe my trajectory transformation is not so accurate but I think that the problem is not that, the problem is that, from my observations, evaluation criterion is easier to achieve the biggest the keypoint size be.
Thanks and greetings
Víctor Mondéjar