ORB descriptor

Source: http://medium.com/data-breach/introduction-to-orb-oriented-fast-and-rotated-brief-4220e8ec40cf Backlinks: Descriptors in feature detection/extraction

• Oriented FAST and Rotated BRIEF, developed 2011
• Was developed as an alternative to SIFT and SURF, and ended up being better/faster than both
• Build on
  • FAST keypoint detector
  • BRIEF descriptor

ORB using FAST, but with (partial) scale invariance

• Use a multiscale image pyramid

• Each level of the pyramid is the same image, but scaled at different resolutions (reduced size as you go higher up)

• Once the pyramid is computed, FAST is used to detect keypoints 

ORB detection

• Compute scale pyramid
• Use FAST to detect keypoints using the scale pyramid
• An orientation is assigned to each keypoint (left/right facing)
  • orientation depends on how the intensity changes around the keypoint
  • intensity detection: intensity centroid

Brief (binary robust independent elementary feature)

• Converts all keypoints into a binary feature vector, in order to be able to represent an object with this vector
• Binary feature vector = binary feature descriptor
• Each keypoint is assigned a feature vector (128 to 512 bit-string)

How it works

• Image smoothing (Gaussian blur) to reduce descriptor’s sensitivity to high-frequency noise
• For one bit within the feature vector, perform a binary test:
  • select a random pair of pixels in a neighbourhood (patch) around the keypoint
    • p1 drawn from a Gaussian distribution centred around the KP, std = sigma
    • p2 drawn from a Gaussian distribution centred around p1, std = 2sigma
  • If p1 brighter than p2, bit is set to 1, else 0
• If 128-bit vector, this process is repeated 128 times, with 128 random pairs

tau: binary test p: patch p(x): intensity at x

Resulting binary vector of one feature (one keypoint) using n binary tests

Drawbacks BRIEF isn’t invariant to rotation –> rBRIEF (rotation-aware BRIEF)

rBRIEF For one feature with n binary tests we have the matrix S

ORB steers BRIEF according to the keypoint orientation, via With theta = patch orientation, Rtheta = rotation matrix

Resulting rBRIEF feature vector