How to extract ORB features from patches

[nkwangyh]

Hi, I am trying to re-evaluate ORB on the patches benchmark and I have extracted ORB feature based on the following rules:

1. take the center of a 65x65 patch as the position of a keypoint
2. compute the angle of the keypoint according to ICAngles from OpenCV
3. set all other attributes of a cv2.KeyPoint such as response, octave, size, class_id to 0 and build a cv2.KeyPoint
4. compute ORB descriptor via cv2.ORB.compute for the KeyPoint

As a result, I achieved a visibly lower performance than that in paper at about 3.x in the image matching task. So I further tried the following settings:

1. resize patch from 65x65 to 31x31
2. extract keypoints and descriptors from the patch image directly without setting patch centers as keypoint positions

But with 1 I still got a similar score at 3.x and with 2 I found the descriptor bits were significantly different from that you offered in the pre-computed descriptor files for the same patch.
So my question is whether my operations are correct, and if not, could you give me some suggestions? Thanks in advance!