I have a use case where I need Akaze to be flip invariant. I have seen this being done for SIFT in the paper “MAX-SIFT: FLIPPING INVARIANT DESCRIPTORS FOR WEB LOGO SEARCH” but I am locked into using Akaze for my application. I assume that flipping the image changes the order of elements in the (MLDB-Upright) descriptor? However, after a lot reading of the Akaze (opencv version) source code/playing around with test data I have not figured it out. Any ideas for figuring out the re-order??
What about flipping the image and extracting AKAZE descriptors?
Descriptors tend to be flipping variant by design, because in the general use case two mirrored descriptors must be different.
Modifying AKAZE descriptors can be a hard work, and if you do you should publish a paper. This is AKAZE paper. Usually papers don’t say much about implementation. There’s no public AKAZE reference implementation, but I seem to remember that author Pablo Alcantarilla worked on the OpenCV implementation.