[ptx] Registration testcases

Bruno Postle bruno at postle.net
Wed Jan 21 10:12:53 GMT 2004


On Tue 20-Jan-2004 at 07:27:53AM +0100, Pablo d'Angelo wrote:
> 
> the examples have been matched using sift features. I'm interested how the
> triangle approach works, compared to these, but my implementation is still
> broken.

I tried just six of the photos from one of the sets, it would take
far too long to try and match them all with the current
implementation of the triangle/distance matching perl-code:

    http://bugbear.blackfish.org.uk/~bruno/distance/preview3.jpg

This is with 50 points per picture, note that two of the photos
were not matched at all (which I suppose is better than a mismatch).

I still have to lens-correct the pictures before identifying
features; I thought it would be ok simply reversing the lens
correction equation and applying that to the coordinates, but this
isn't accurate enough.

I have some ideas about speeding-up up the algorithm, though I think
the ultimate flaw with this process it that you have to know the
hfov, a, b & c lens parameters precisely before you start.

-- 
Bruno


More information about the ptX mailing list