This paper investigates how phase correlation can be used to make a robust feature descriptor. It was later replaced by the LPM as a more efficient descriptor. Nonetheless, it introduces some important things like the fact that many implementations of the log-polar transform is not fit to be used for small images like feature patches as it does not sample from the centre. The image to the left is sampled in a log-polar fashion, with its centre in the small white dot in the central black pixel. The upper left image shows the result of using OpenCV, where the white dot appears like a dot and not a long white band as it should (as is seen on our method on the lower right). We tried many other Matlab implementations that we found on the web and some of them exhibited exactly the same behaviour, while others did not show the white pixel at all. We also sample on a disk rather than on a square to make the patch rotation invariant, using a 90 degree rotated filling order, which explains the difference in the right images.

The other thing this paper investigates is to use a Gaussian to be able to sample “between” pixels, by averaging together neighbouring pixels using the Gaussian.

  • Rotation Invariant Feature Matching – Based on Gaussian Filtered Log Polar Transform and Phase Correlation. A. Hast, A. Marchetti. ISPA 2013, Full Paper. pp. 100-105. 2013.
    @inproceedings{Has13a, author = {A. Hast and A. Marchetti}, title = {Rotation Invariant Feature Matching - Based on Gaussian Filtered Log Polar Transform and Phase Correlation}, booktitle = {ISPA 2013}, pages = {100--105}, year = {2013}}


Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s