[Computer Vision 6] ORB: an efficient alternative to SIFT or SURF

ORB: Oriented FAST keypoints and Rotated BRIEF descriptors

Related Works

FAST keypoints:  Using Harris corner filter to find corner keypoints. SIFT and SURF can achieve orientation variant but computational, while centroid operator gives a single (efficient) orientation for keypoint.

BRIEF descriptor: very sensitive to in-plane rotation.  To look for the test is least sensitive to rotation, we need to remove uncorrelated information.

FAST Keypoint Orientation

FAST Parameter: the intensity threshold between centre and a circular ring about the centre. 

Add a scale pyramid to achieve scale invariant, apply a Harris Detector to detect corners in each layer of the pyramid (set a low threshold to obtain lots of points, then order them in Harris measure, and pick the top scored points).

Intensity Centroid: A corner's intensity is offset from its centre. We find the orientation of the vector OC. O is the corner's centre, C is the intensity centroid.

rBRIEF: Rotation Aware BRIEF

BRIEF: sum of n binary test of n sets of points (X, Y) inside a patch, binary test compare the intensity of point X and point Y.

To achieve rotation invariant, we steer the BRIEF according to the orientation of keypoint. By multiplying a rotation matrix R, we obtain new sets of points.