sift extract with less keypoints for megadepth datasets

[AbyssGaze]

Thanks for your amazing work. I have implemented superpoint extact superglue on megadepth with some imperfection(depth loss when interpolate, I will discard that image pair). The submission in https://www.visuallocalization.net/ is : My implementation with online keypoints extraction, if superpoints less than 1024/2048 I will random select points with interpolation descriptors, but when migrating to sift, some images with too little keypoints, how to deal that problem? Thanks

[sarlinpe]

Great job, this looks pretty good! For SIFT we randomly sampled points but also descriptors, as the absolute value of a Gaussian in a D256 space, potentially with root normalization to emulate RootSIFT.

[AbyssGaze]

Thanks awfully! I will try it and public RootSIFT results.

[zenmedou]

Thanks for your amazing work. I have implemented superpoint extact superglue on megadepth with some imperfection(depth loss when interpolate, I will discard that image pair). The submission in https://www.visuallocalization.net/ is : My implementation with online keypoints extraction, if superpoints less than 1024/2048 I will random select points with interpolation descriptors, but when migrating to sift, some images with too little keypoints, how to deal that problem? Thanks

Hi，this is really high performance! May I ask if it is combined with hloc or not? Thanks.

[AbyssGaze]

Yes, It's combined with hloc. I just have reproduced Sarlin's work(superglue with megadepth).

[AbyssGaze]

@Skydes Thanks for your suggestion, I have used rootsift. Although the loss decreases and the matching performance get better, the loss decreases to 1.0 and it converges. Does sift(DOG) keypoints need to do nms as superpoint? Thanks!

[zenmedou]

HI, thanks for the reply. May I ask what if hloc is not added? I'm also trying to reproduce results on megadepth but have some performance gap for sg+sp. Approximately 77 vs 79(official model) on aachen. I'm using offline data generation which yields about 600k pairs in total.

[sarlinpe]

@AbyssGaze DoG does not need any NMS but we you need to define tighter thresholds when computing the ground truth, as keypoints are generally better localized. The exact value depends on your image resolution.

[szgy66]

@Skydes Thanks for your suggestion, I have used rootsift. Although the loss decreases and the matching performance get better, the loss decreases to 1.0 and it converges. Does sift(DOG) keypoints need to do nms as superpoint? Thanks!

hello,I want to ask that how combline sift with superglue? The length of the SIFT descriptor is 128D, but the superglue input is 256D? And does Superglue need to retrain? thanks!

[AbyssGaze]

Yes, we will publish the training codes no surprise on May.1.2022.

Yang Gao @.***> 于2022年1月22日周六 16:19写道：

@Skydes https://github.com/Skydes Thanks for your suggestion, I have used rootsift. Although the loss decreases and the matching performance get better, the loss decreases to 1.0 and it converges. Does sift(DOG) keypoints need to do nms as superpoint? Thanks! [image: 14_400_matches] https://user-images.githubusercontent.com/14217667/112408523-e7c1bb00-8d52-11eb-9d1c-8d7b4363ba17.png

hello,I want to ask that how combline sift with superglue? The length of the SIFT descriptor is 128D, but the superglue input is 256D? And does Superglue need to retrain? thanks!

— Reply to this email directly, view it on GitHub https://github.com/magicleap/SuperGluePretrainedNetwork/issues/70#issuecomment-1019097499, or unsubscribe https://github.com/notifications/unsubscribe-auth/ADMPDQ6AEJWCTEI4V3M25XDUXJR7ZANCNFSM4ZGPQXPQ . Triage notifications on the go with GitHub Mobile for iOS https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675 or Android https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub.

You are receiving this because you were mentioned.Message ID: @.***>

[yangfuwei]

Hi, 1-May-2022 have passed, Did you publish your training code as planned?

Yes, we will publish the training codes no surprise on May.1.2022. Yang Gao @.> 于2022年1月22日周六 16:19写道： … @Skydes https://github.com/Skydes Thanks for your suggestion, I have used rootsift. Although the loss decreases and the matching performance get better, the loss decreases to 1.0 and it converges. Does sift(DOG) keypoints need to do nms as superpoint? Thanks! [image: 14_400_matches] https://user-images.githubusercontent.com/14217667/112408523-e7c1bb00-8d52-11eb-9d1c-8d7b4363ba17.png hello,I want to ask that how combline sift with superglue? The length of the SIFT descriptor is 128D, but the superglue input is 256D? And does Superglue need to retrain? thanks! — Reply to this email directly, view it on GitHub <#70 (comment)>, or unsubscribe https://github.com/notifications/unsubscribe-auth/ADMPDQ6AEJWCTEI4V3M25XDUXJR7ZANCNFSM4ZGPQXPQ . Triage notifications on the go with GitHub Mobile for iOS https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675 or Android https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub. You are receiving this because you were mentioned.Message ID: @.>

[AbyssGaze]

I'm very sorry, and now our open source process has become very complicated and demanding. We need to complete the OETR open-source process first, and we need to go through the magic of legal affairs. Suppose there are more papers published in the future. In that case, we will integrate them into the SuperRegistrator codebase (including code integration libraries such as superglue, bihome, and loftr) before starting the open-source process again. I am very sorry.