Different results when trying MKD, TFeat descriptors in kornia

[jac08h]

Hello,

I tried to run MKD_TFeat_descriptors_in_kornia.ipynb, but my results differ significantly from the example notebook. Most importantly, using deep-learning-based descriptors does not yield better matches.

I did a single change to the example code: Instead of matchesMask = inliers_mask.ravel().tolist(), I use inliers_mask.astype(int).ravel().tolist(), because the former was throwing a cv2.error - "Can't parse 'matchesMask'. Sequence item with index 0 has a wrong type".
Package versions:

• kornia: 0.6.10.dev0
• cv2: 4.7.0
• pydegensac: 0.1.2
• kornia_moons: 0.2.3

The output is in this notebook.

Do you have an idea what is causing the different behaviour?
Thanks a lot for your time!

[ducha-aiki]

Thanks for the report. The problem is that OpenCV drawing code stopped working at some point.

inliers_mask.astype(int).ravel().tolist()

is wrong. when you convert bool (the correspondence is true or not) to int, you instead of filtering, index the points.
inliers_mask.astype(bool).ravel().tolist() is correct in principle, but you have to play with OpenCV to make it work.

Or - better - switch to kornia_moons visualization, as in this tutorial:

https://kornia-tutorials.readthedocs.io/en/latest/image_matching_adalam.html

[jac08h]

Thank you for the response.

The example with KeyNet-AdaLAM works for me - I get the same visualization and also the same amount of tentative matches and inliers.

However, using kornia_moons visualization does not fix the problem with MKD and TFeat descriptors - it displays the same matches:

Therefore, I believe the problem is in the matches/descriptors themselves. For example, fewer inliers are found in my run.

[ducha-aiki]

@jac08h could you please share updated colab?

[jac08h]

Here is the version using kornia_moons visualization. To use it, I modified get_local_descriptors to return LAFs as well.

[ducha-aiki]

@jac08h I don't know why, but you are using very old version of kornia-examples.

The key difference is in current version we have:

timg = K.color.rgb_to_grayscale(K.image_to_tensor(img, False).float()/255.)

Whereas in yours:

timg = K.color.rgb_to_grayscale(K.image_to_tensor(img, False)/255.)

We have dropped auto-conversion in K.image_to_tensor in kornia to float long time ago, so when image is divided by 255, it essentially becomes binary image.

After the fix above:

[jac08h]

You are right, that was the issue. Thank you so much for your quick help.