search

rpautrat / SuperPoint

Efficient neural feature detector and descriptor
MIT License
1.88k stars 416 forks source link

Do you have plan to improve hpatches-v? #241

Closed litingsjj closed 2 years ago

litingsjj commented 2 years ago

Sorry to bother you. this is a great project! I have a question about hpatches-v result. do you have any plan or advices to improve hpatches-v result for superpoint?

rpautrat commented 2 years ago

Hi, I am not currently working on the improvement of SuperPoint. But since the publication of SuperPoint, there have been multiple works that improved the performance of the detector and descriptor. For example for points, you could check KeyNet and DISK.

litingsjj commented 2 years ago

Thanks a lot! I compared some images with superpoint and sift. For scale change, sift is much better than superpoint(or other cnn methods). the result is so terrible if the object in image1 is four times than in image2. I will read KeyNet and DISK that you provided. Also for descriptor matches, did you have some advices to improve? There are some paper used transformer, like superglue, LoFTR to get a better result, but the speed is so slow. I will so appreciate if you can solve my confusion, even if provide some clues. Thanks again!

rpautrat commented 2 years ago

Yes, learned descriptors have notoriously problems with scale changes, and SIFT and still very good for that. This is still an active area of research, so I can't help you much on that. The recent learned machers like SuperGlue and LOFTR are indeed very good, but slower. There will always be some trade-off, it depends on your application.

litingsjj commented 2 years ago

thanks for your reply!