findHomography vs. pnpsolver
In my image recognition software I use ORB features and "findHomography" with RANSAC to validate my initial feature matches made by a Flannindex. With OpenCV 3 I consider to switch to AKAZE features. Unfortunately my tests with AKAZE features point out, that the time consumption of queries with AKAZE features is al lot higher, than the time consumption when using ORB. Furthermore the matching results are notably worse. With ORB no false positive matches remain after the reranking with "findHomography" with RANSAC. AKAZE produces a lot false positives matches and the number of true positives matches is lower than number of true positives matches of ORB. Does anyone have experience with AKAZE features in combination with "findHomopgrahy?
Additionally is made some experiments with the PnP Ransac solver of OpenCV 3. But the results are not good, too. Is it sufficient to use the default parameters in this case?
May be you can find some usefull results here
Shouldn't findHolography be independent of which features descriptors you actually use? Why would it work worse with Akaze than with Orb when Akaze has much more matches then Orb normally?
yes findHomography does not depend of features descriptors. But it depends of matches process. If you have got many matches and many are wrong that's a problem even with ransac.
If you want to improve findHomography you can sort your match as function of distance. you can try to match image A with B and B image with A and keep only match which are symetrics