Thread by @ducha_aiki

While I am selecting nice pictures for my PhD thesis, let's do a fun visual benchmark of local features on the photos I took myself.

SIFT-HardNet, R2D2, SuperPoint, SuperGlue
1/

Nothing works here, ha :)

SIFT-HardNet, R2D2, SuperPoint, SuperGlue
2/

SIFT-HardNet, R2D2, SuperPoint, SuperGlue

As in 1, only SuperGlue works :)
3/

https://twitter.com/ducha_aiki/status/1330495990772740106/photo/1

SIFT-HardNet, R2D2, SuperPoint, SuperGlue

SIFT-HardNet sucks, R2D2 and SP kind of work, SG rocks
4/

R2D2 here hallucinates a lot of false correspondences, while SuperPoint works quite well even without SuperGlue (which rocks, as usual)
5/

Here, all 3 are comparable. 6/

HardRock cafe in Prague (EVD dataset)
It looks like everything works, but actually repeated patterns are matched incorrectly (including some of SG) 7/

SIFT-HardNet. R2D2, SuperPoint, SuperGlue
SuperPoint is the only detector, which manages to match something on the monument, not the plate only
Aaaand, somehow SuperGlue does not like the plate, so it kills all the correspondences there
8/

R2D2 fails here, others are fine
9/

SIFT-HardNet and R2D2 fail here, SuperPoint rocks.
10/10
Overall, for random unknown upright pair of images with possible illumination change, I would go for SuperPoint, as least if we don't consider affine view synthesis.

11/10
Here DISK results. DISK is quite cool
@jatentaki

DISK, continue.
The following pairs: stadium, ministry (with river), dancing house -- no correspondences (with 8k detections).

@jatentaki while you are here, could you please take a look at the PR? :)
github.com/cvlab-epfl/dis…