Re: [libcvd-members] Problems with fast corner code

[Edward Rosten]

On Fri, 3 Aug 2007, Georg Klein wrote:

> This is with fast_corner_detect_10, but probably applies to others as well.
>
> fast_corner_detect_10, when using faster_blah_10, crashes on images whose 
> width is not a multiple of 16; this is because of the lines
>
>         const byte* const end = I[I.size().y - 3]-16;
>         for (const byte* p = I[3] + 16; p!=end; p+=16) {
>                                        ^^^^
>
> For images where size isn't a multiple of 16 but the check still works (e.g. 
> width multiple of 8, even number of rows) the code doesn't crash, but 
> produces spurious corners at the image edges.
>
> Before the recent fast merge, this was not a problem since I was using slow 
> FAST for the cases where this comes up. I don't understand the faster code 
> enough to fix this myself, alas.

OK, This is not my code, but I'll do my best here...

The code runs through the image testing 16 pixels at a time for FAST-ness. 
It doesn't bother going row by row, it just treats the image as a large 
contiguous block and rejects the corners detected too close to the edge 
in a post processing step.

It also checks the first few and last few pixels "by hand" since it can't 
use SSE for them.

faster_9 is the same and it probably suffers from the same bug.


> What I could do is add checks to the logic 
> which decides which detector to use, but I don't think that's the right way, 
> no?

I've checked in some new code, which may or may not contain a fix, but you 
can test it if you wish. This one instead goes row by row, but the logic 
is simpler. As a bonus, it's also "sub_image ready", in that it is a 
trivial change to make it work on subimages.

Please let me know if this works because I think the other detectors will 
share the bug, and they need to be fixed.


I don't know if I've really wrecked the speed of it with my changes. I 
expect not too much, since my recent experience suggests that the compiler 
is exceptionally good at "doing the right thing" when indexing images, and 
[] based code can be faster than pointer based code.

-Ed