fastApprox match always gives same feedback as bruteForce() match about k2nn HOT 2 CLOSED

Hmm. tho in another case fast-approx gives no results whilst brute-force gives the kind of results im expecting.. so something i'm not understanding here

from k2nn.

Hi,

I don't currently have time to examine external code-bases, but I can say that fastApproxMatch is an experimental attempt to do just a single twiddle pass of multi-index hashing. It will only return guaranteed matches, but it will only find such matches if the query and training differ by less than (32 - threshold), so it could work well in a high-frame-rate VO application where each frame differs very little, but is not intended, and not likely to work, for a SfM application like yours, where thresholds must be high as matches will still differ a lot.

I suggest sticking to bruteMatch for your application.

If even bruteMatch causes problems, check things like degrees and radians. Do a simple test and confirm the pipeline is working on two images, like I did with this gentleman for CLATCH:
komrad36/CLATCH#1

That way you can test rotation invariance, scale invariance, etc.

Also note that the CPU version necessarily handles scale sort of imperfectly; as I think I mentioned previously somewhere, it's great at handling various scales, i.e. at matching small scale features, or matching large scale features, but not good at matching small scales to large ones.

The CUDA version is a little better, but not great.

A really good solution is planned by tightly coupling the upcoming KORB to CLATCH such that they generate and share pre-computed interpolated scale spaces. That'll work great. Probably take a couple more weeks.

Thanks,
Kareem

from k2nn.