vl_covdet_extract_laplacian_scales_for_frame uses the wrong peak threshold about vlfeat HOT 5 CLOSED

Thank you for reporting this problem.

self->peakThreshold is relative to the cornerness measure. This would coincide with the Laplacian scale score for a Laplacian-based detector, but would differ for Harris corner points or Hessian. Thus using self->peakThrehsold would not be correct for Harris-Laplace or Hesisan-Laplace.

The correct solution is to add a new parameter to the covariant detector class covdet to specify this parameter independently as an additional threshold.

from vlfeat.

Sure, that sounds good too. Unfortunately, the scale of my data is very different than typical so this threshold is highly relevant for me. Thanks.

from vlfeat.

I implemented a patch in the devel branch. Could you check it out to see if that address your problem?

from vlfeat.

I'm not seeing it in the list of commits, has it been pushed to github?

from vlfeat.

Ah, ok, now I see it. I've taken a look and that should fix it, thanks!

from vlfeat.