One of the properties of BRIEF is that each bit feature has a large variance and a mean near 0.5. A mean of 0.5 gives a maximum sample variance of 0.25 for a bit feature. Steered BRIEF produces a more uniform appearance to binary tests. High variance causes a feature to respond more differently to inputs.
Having uncorrelated features is desirable as in that case, each test has a contribution to the results. We search among all the possible binary tests to find ones that have a high variance (and a mean close to 0.5) as well as being uncorrelated.
ORB specifies the rBRIEF algorithm as follows.
Set up a training set of some 300 keypoints drawn from images in the PASCAL 2006 set. Then, enumerate all the possible binary ...