Ask for advice about merging match_keypoints into hloc sfm pipeline

[DabblerGISer]

I've tried using superpoint and also dedode as keypoints extractor and then match with RoMa but can't get a proper result. How could I improve the result?
The basic idea is that I firstly extract all keypoints and then normalize the keypoints into range [-1, 1].

def to_reference_coord(kpt, W, H):
kpts = torch.stack((kpt[..., 0]/(W/2) -1, kpt[..., 1]/(H/2) -1), axis=-1)
return kpts

And then using roma model to calculate the warp and certainty. then using kpts, warp and certainty as input of match_keypoints. After getting the results I also used ransac to filter some of the matches.

I'm not able to reconstruct the model correctly by this way.

Thus I check the visual result of matches:

Anybody could give me some advice to improve the pipeline? Thanks!

[DabblerGISer]

By the way, this magic pickup truck is really hard to reconstruct. LOL

[Parskatt]

I would avoid matching when the roma confidence is very low. In the provided images I would guess that is the case.

[DabblerGISer]

I would avoid matching when the roma confidence is very low. In the provided images I would guess that is the case.

Actually RoMa reaches pretty accurate result by RegressionMatcher.match. Thank you for your reply, I will firstly have a review on my code.