Paper Review: FREAK: Fast Retina Keypoint


Authors: Alexandre Alahi, Raphael Ortiz, Pierre Vandergheynst


Q1: What is the formula for the retina pattern?

The one difference from BRISK is the pattern has overlapping circles. In BRISK they were tangential. Redundancy increases recognition.

The circles are log polar. In this case, it's similar to Shape Context descriptors, but we don't divide into regions, we create increasingly larger circles on polar lines.

Q2: What do the descriptors contain?

Binary descriptor is a string of bits. A bit corresponds to a pair of receptive fields. If intensity of the first receptive field is larger than the second, bit is set to 1, otherwise it's zero.

Q3: How the sampling works?

Each circle in the pattern is called receptive field. Gaussian kernel is applied to these fields and their intensity are calculated.

Q4: Is there scale invariance? How?

There is no discussion of scale invariance, but scale invariance seems to arise from the building of the descriptor. Since the circles are created in log polar orbits and bits are put into descriptor according to their contribution to recognition, scale invariance follows these.

Q5: How the rotation invariance achieved?

Orientation is calculated using 45 symmetric pairs from the center. It has larger steps than that of BRISK and thus needs lower memory.