![](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F210302%2Fe0f4fbda6256645275eb20b40fe242f9%2FScreenshot%202024-04-03%20at%2011.14.09.png?generation=1712135667888200&alt=media\")
Given the threads 1 and 2 about license and SuperGlue in particular, we decided to address also a performance question.
\nWe understand, that given the dominance of SuperGlue in Image Matching Challenge since 2020 until 2023, it is kind of hard to believe that not using it is a good idea.
\nHowever, license issues aside, LightGlue is better than SuperGlue performance-wise as well, so there is no good argument to keep using SuperGlue.
Specifically, let me point out results on IMC-2021 leaderboard, especially with ALIKED features. It provides better result than SuperGlue in all entries, except those, which use additional features and pre/post-processing.
\nYou should check the \"MultiView\" score in particular
https://www.cs.ubc.ca/research/image-matching-challenge/2021/leaderboard/
\n
![](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F210302%2F24a32d444e487a5adabf509b494d7c74%2FScreenshot%202024-04-03%20at%2011.15.21.png?generation=1712135737534378&alt=media\")
![](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F210302%2Ff6c4fa1d2dd61b945b30d91281d1680c%2FScreenshot%202024-04-03%20at%2011.16.07.png?generation=1712135783594914&alt=media\")
Besides IMC-2021 dataset (and all the comparisons in the LightGlue paper itself), here is a comparison on WxBS and EVD datasets
\nhttps://github.com/cvg/glue-factory/pull/52#issue-2065776066
![](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F210302%2Fdbb3b871c83cfa7ee654a85aa5a3e966%2FScreenshot%202024-04-03%20at%2011.17.05.png?generation=1712135840436199&alt=media\")
One particular technical issue with LightGlue is that its default parameters are optimized for speed, not maximum performance, so it might seem to be worse than SuperGlue.
\nHowever, you can turn off all the speed optimizations with setting up depth_confidence and width_confidence to -1:
matcher = LightGlue(=, =-1, =-1)\nI hope, that this thread, together with license part of the story, will close the SuperGlue issues.
", "messages": [ { "id": 2732704, "postDate": "2024-04-03T09:29:05.283Z", "content": "Given the threads 1 and 2 about license and SuperGlue in particular, we decided to address also a performance question.
\nWe understand, that given the dominance of SuperGlue in Image Matching Challenge since 2020 until 2023, it is kind of hard to believe that not using it is a good idea.
\nHowever, license issues aside, LightGlue is better than SuperGlue performance-wise as well, so there is no good argument to keep using SuperGlue.
Specifically, let me point out results on IMC-2021 leaderboard, especially with ALIKED features. It provides better result than SuperGlue in all entries, except those, which use additional features and pre/post-processing.
\nYou should check the \"MultiView\" score in particular
https://www.cs.ubc.ca/research/image-matching-challenge/2021/leaderboard/
\n
Besides IMC-2021 dataset (and all the comparisons in the LightGlue paper itself), here is a comparison on WxBS and EVD datasets
\nhttps://github.com/cvg/glue-factory/pull/52#issue-2065776066
One particular technical issue with LightGlue is that its default parameters are optimized for speed, not maximum performance, so it might seem to be worse than SuperGlue.
\nHowever, you can turn off all the speed optimizations with setting up depth_confidence and width_confidence to -1:
matcher = LightGlue(=, =-1, =-1)\nI hope, that this thread, together with license part of the story, will close the SuperGlue issues.
", "rawMarkdown": "Given the threads [1](https://www.kaggle.com/competitions/image-matching-challenge-2024/discussion/490622) and [2](https://www.kaggle.com/competitions/image-matching-challenge-2024/discussion/490714) about license and SuperGlue in particular, we decided to address also a performance question.\n\nWe understand, that given the dominance of SuperGlue in Image Matching Challenge since 2020 until 2023, it is kind of hard to believe that not using it is a good idea. \nHowever, license issues aside, LightGlue is better than SuperGlue performance-wise as well, so there is no good argument to keep using SuperGlue.\n\nSpecifically, let me point out results on IMC-2021 leaderboard, especially with ALIKED features. It provides better result than SuperGlue in all entries, except those, which use additional features and pre/post-processing.\nYou should check the \"MultiView\" score in particular\n\nhttps://www.cs.ubc.ca/research/image-matching-challenge/2021/leaderboard/\n\n\n\n\n\n\n\n\n\nBesides IMC-2021 dataset (and all the comparisons in the LightGlue paper itself), here is a comparison on WxBS and EVD datasets\nhttps://github.com/cvg/glue-factory/pull/52#issue-2065776066\n\n\n\n\nOne particular technical issue with LightGlue is that its _default_ parameters are optimized for speed, not maximum performance, so it might seem to be worse than SuperGlue. \nHowever, you can turn off all the speed optimizations with setting up `depth_confidence` and `width_confidence` to `-1`:\n\n```\nmatcher = LightGlue(features='FEATURENAME', depth_confidence=-1, width_confidence=-1)\n```\n\nI hope, that this thread, together with license part of the story, will close the SuperGlue issues.", "votes": 27 } ], "comments": [], "raw_markdown_by_id": { "2732704": "Given the threads [1](https://www.kaggle.com/competitions/image-matching-challenge-2024/discussion/490622) and [2](https://www.kaggle.com/competitions/image-matching-challenge-2024/discussion/490714) about license and SuperGlue in particular, we decided to address also a performance question.\n\nWe understand, that given the dominance of SuperGlue in Image Matching Challenge since 2020 until 2023, it is kind of hard to believe that not using it is a good idea. \nHowever, license issues aside, LightGlue is better than SuperGlue performance-wise as well, so there is no good argument to keep using SuperGlue.\n\nSpecifically, let me point out results on IMC-2021 leaderboard, especially with ALIKED features. It provides better result than SuperGlue in all entries, except those, which use additional features and pre/post-processing.\nYou should check the \"MultiView\" score in particular\n\nhttps://www.cs.ubc.ca/research/image-matching-challenge/2021/leaderboard/\n\n\n\n\n\n\n\n\n\nBesides IMC-2021 dataset (and all the comparisons in the LightGlue paper itself), here is a comparison on WxBS and EVD datasets\nhttps://github.com/cvg/glue-factory/pull/52#issue-2065776066\n\n\n\n\nOne particular technical issue with LightGlue is that its _default_ parameters are optimized for speed, not maximum performance, so it might seem to be worse than SuperGlue. \nHowever, you can turn off all the speed optimizations with setting up `depth_confidence` and `width_confidence` to `-1`:\n\n```\nmatcher = LightGlue(features='FEATURENAME', depth_confidence=-1, width_confidence=-1)\n```\n\nI hope, that this thread, together with license part of the story, will close the SuperGlue issues." } }