Re: [libcvd-members] FAST corner detection--- RFC

[Georg Klein]

On Mon, 11 Jun 2007, Edward Rosten wrote:

> > > Does anyone ever use FAST corners without nonmax? I think that nonmax 
> > > suppression should be the default operation. Is there a good reason to 
> > > provide non nonmax functions?
> >
> > Yes, I do. I initialise my features on max locations, then run detection on 
> > all cornery regions.
>
> Interesting.
>
> How about having
>
> fast_corner_detect_XX()
>
> and
>
> fast_corner_detect_XX_nonmax()

Sure, but it'll break everyone's existing code :)
The slightly uglier but more compatible alternative is to have a different
function name which does all-in-one.
It's up to you IMO. The code breakage would be trivial to fix.


By the way - am I correct in my vague recollection that the nonmax 
function currently doesn't even know about the xx in fast-xx? I suppose 
your new formulation would fix that?

// wish list start
If your new nonmax-suppression code will fix that, can you make standalone 
versions of that as well? And if it's faster, maybe even a version which 
spits out both lists at once? Finally, a version which outputs 
corners-and-scores would allow better dynamic thresholding.
// wish list end

> How so? What machines does it fail on?

It didn't on mine here for a while (core 2 duo) but now does. Any number 
of things could have changed.

> > that you are in fact getting the SCREAMING SIMD version if you want it. 
> > Currently, the compile brekage if you don't accomplishes that.
>
> How so? It's supposed to use non SSE versions if it doesn't detect it.

Was probably something broken on my end - I was getting link errors. Maybe 
.so and .h out of sync for some reason, although that shouldn't be 
possible.