Detector repeatability study

[Pierre Moulon]

Hello,

You will find in attachment a repeatability study over the libmv keypoints detectors.

Objective :

 Test the repeatability of FAST/STAR/SURF libmv keypoint detector.

Cheers,

I'm waiting your comments ;)

Pierre 

Here a resume of my conclusions :

We have now an executable main (that I will commit soon) and a "test set" to test our detector over a camera ground truth.

We have seen that our detectors are not so bad and that FAST and STAR seems to be the winner in term of repeatability and position accuracy.

It’s clear that the next step is to consider rotation invariance for FAST/STAR/SURF detector to have a complete Keypoint detector system. Subpixel estimation
could be a plus too.

A comparative study to state of the art could be a plus (Plug SIFT/ others) in my code will be possible (via an executable calling)... and so give a better overview of our performance...

TODO and future work:
Now we must take different task in option.
  - Test the accuracy of scale estimation and robustness to rotation. For the moment STAR and SURF are scale invariant

  - Test the accuracy of the rotation estimation when we will add it.
  - Consider subpixel estimation to achieve best position accuracy in those test set.
  - Consider to compute a minimal subset of keypoint on image rather than export the entire set of computed feature. I.E try to keep the 800 Keypoints
that get the better score...
- We must consider other test set to consider pure rotation, scale, cropping.

- I want also add the test set build for ASIFT http://mw.cmla.ens-cachan.fr/download/megawave-algo/asift/dataset_Morel_Yu_09.zip that could be an excellent addition to the one from Mikolajczyk ..

- Make the same work for descriptor repeatability and accuracy...