Search code examples
computer-visionorb

Explain about rBRIEF and steered BRIEF in ORB

I am reading original paper about ORB. I find it really confusing when reading to descriptors in rBRIEF. Here the paper says

We also enumerate all possible binary tests drawn from a 31×31 pixel patch. Each test is a pair of 5×5 sub-windows of the patch. If we note the width of our patch as wp = 31 and the width of the test sub-window as wt = 5, then we have N = (wp − wt)^2 possible sub-windows. We would like to select pairs of two from these, so we have N 2 binary tests. We eliminate tests that overlap, so we end up with M = 205590 possible tests

Before rBRIEF, the author mentioned steered BRIEF, and it uses orientation to construct new distributions of (x,y) to find the binary. But in rBRIEF, I don't see they mention about orientation, the just said that 'Each test is a pair of 5x5 sub-window', so what is the point in that 5x5 sub-window needed to get to compare intensity with another point in another 5x5?

If you find it hard to understand what I write, the question is: Which is the point in 5x5 sub-window I need to find to compare intensity with another point in another 5x5 sub-window? Is orientation contributes to rBIREF ? And if yes, How?

Thank you so much

Solution

  • Which is the point in 5x5 sub-window I need to find to compare intensity with another point in another 5x5 sub-window?

    center point of that window (which is also why this window size and convolution filters in general are odd). such a window is used because smoothing is applied before the actual pixel (intensity) comparison. making it more robust to noise while having the drawback of losing some information

    Is orientation contributes to rBIREF ? And if yes, How?

    sBRIEF compensates for rotation changes by using an arbitrary detectors (e.g. SURF) approximation of the angle of a patch

    rBRIEF is the ORB authors version of compensating for rotation changes, as it is based on BRIEF descriptor (in short they use image moments to calculate the centroid of the patch and do the rotation while using a learning strategy to overcome the shortcomings of sBrief, ie. loss of variance)