Issue with homography estimation

[thereza]

I've noticed that some components map very well while others do not after applying the homography estimation. Wondering what you used to generate your library. One issue is that homography estimations typically occuring using 6DOF which takes perspective into account (i.e. 3d rotation). As we are operating only in 2D, we need a 4DOF estimation else there is a possibility of introducing error. check out the first answer here - http://stackoverflow.com/questions/2250 ... -in-opencv - not sure if it applies to the code you are using.

[JuKu]

I bought it, building that from scratch is beyond my image processing capabilities. can you elaborate on the issues you find? Especially, if you have a bad behaving board, I'd like to have a look (and possibly, get the transform programmer involved).

[thereza]

Some points map well, but others don't.

In the source, you mention being able to measure the skew, etc -- the skew should be 0 and if it's anything but zero then there is an error. The fact that the code enables that measurement means that it's not detecting a rigid transformation, and by introducing additional degrees of freedom, it might think the board moved in 3d space which causes the distortion.

At least that's my suspicion.

I'm looking at implementing this approach using the opencv libs (for solving SVD) (already implemented but i dont trust it and need to test) - http://nghiaho.com/?page_id=671 as a way of dealing with a rigid transformation.

Reza