Quality findHomography OpenCV 3 vs. OpenCV 2

edit

Hi all,

I'm using OpenCV on a project for aligning multiple overlapping images. In more detail, I'm using the function findHomography as a feature match filter between a single image pair. In other words, I don't use the actual computed homography but only the inlier mask.

All inlier points of every camera pair are then used in conjunction with the levmar library to estimate camera parameters in a bundle adjustment.

Comparing the generated output when simply switching between linking against OpenCV 2.4.8 and OpenCV 3 (Beta and RC1) the results are pretty consistently worse (in terms of alignment) in the OpenCV 3 Version.

I briefly looked at the source code and noticed that there seem to have been substantial changes between the two implementations.

Has anyone else noticed similar results? Am I missing an option that maintains result compatibility or is this something I have to live with (which would mean not using OpenCV 3 going forward, which would be kind of a shame)?

Any input on these questions would be greatly appreciated.

Thanks in advance!