Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Bad homography transformation obtained from superglue matchpoints. Help! #126

Open
SmileyScientist opened this issue Nov 14, 2022 · 4 comments
Open

Bad homography transformation obtained from superglue matchpoints. Help! #126

SmileyScientist opened this issue Nov 14, 2022 · 4 comments

Comments

@SmileyScientist
Copy link

Hi,

I am using SuperPoint and SuperGlue to extract features and match them to compute homography between two images but the matchpoints seem to be performing worse than SIFT-FAST.

The left image is 1080x1088 and right is 1920x1440. They are captured using two different sensors.
frame_1

The matchpoints look good but if I use them to compute homography using cv2.findhomography(), this is what I get
frame_1_warp_homography

I have tried modifying a lot of parameters but the homography I get using the matchpoints is no up to the mark. I am using "outdoor" superglue to accommodate for multiple sensors whose intrinsic matrices are unknown.

Here is the wrap using SIFT and FAST which looks waaaaay better
frame_2_warp_homography

I want to get a decent homography transformation using SuperPoint and SuperGlue. Any idea on what the issue is and how I can fix it?
Note: the goal is NOT to retrain either of the networks.

All help is much appreciated. Thanks!
@sarlinpe
Copy link
Contributor

The correspondences look fine. The issue is likely due to an incorrect use of cv2.findhomography - I've used it in the past without any problem.

@foxkw
Copy link

foxkw commented Nov 29, 2022

The correspondences look fine. The issue is likely due to an incorrect use of cv2.findhomography - I've used it in the past without any problem.

Why are there so few points that can be matched

@SmileyScientist
Copy link
Author

@skydes , have you used them when the image's keypoints in the world space are not planar?
I recently re-learnt that homography will only work when the points are on a planar surface and in my case, I have points being picked from two different planar surfaces.

@foxkw , the threshold for keypoint match is high. I am after high quality matches. Even though I am using a match threshold of 0.95, I still get some false positives. Do you have any experience with this?

@foxkw
Copy link

foxkw commented Dec 7, 2022

NO. but I found that the number of points you detected is also very small, and I also encountered this situation; I don't know whether this is the cause; I found that the number of matches of the superpoint model I trained with superglue after homography transformation is far less than the official result. I wonder if you have encountered and solved this phenomenon

#129

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Assignees
No one assigned
Labels
None yet
Projects
None yet
Milestone
No milestone
Development

No branches or pull requests
3 participants
@SmileyScientist @sarlinpe @foxkw