evaluation problem

Question

evaluation problem

Opened this issue 4 years ago · 9 comments

Sorry to bother you again,I follow you steps to verify other models,such as R2D2.however,the Homography accuracy obtained is inconsistent with the result of your paper table and there is a big difference between them.please tell me the reason,or could your please tell me specific peocess?thank you very much!

Answer 1 · 2021-01-14T01:20:55.000Z

Please check if the input image is single-scale or multi-scale. R2D2 uses image pyramid as default input, however, the results presented in our paper take a single-scale image as input. The footnote in the paper also presents the results with image pyramid input.

Answer 2 · 2021-01-14T03:19:14.000Z

Thank you for your reply,I changed R2D2 to test under a single scale,but the result is still inconsistent with the result in your paper.could your please write me a specific step?I really need it,and i'm really sorry to bother you and really appreciate your help.

Answer 3 · 2021-01-14T05:12:32.000Z

Please check 3rd_party/r2d2/, I have added the scripts that extract R2D2 features from hpatches dataset. The code is almost the same as the code released by R2D2 authors.

Answer 4 · 2021-01-14T07:43:20.000Z

Thanks for your reply,I finish you step,but my result is still inconsist with the result in your paper,but it's close,my result is: 0.7412068965517241, 0.8891228070175436, 0.5983050847457627
In your paper is 0.7215 0.9375 0.5128.
I also resize the Hpatches dataset,but the answer is different with you,please help,I should have no problem after I finish this.
best wish.

Answer 5 · 2021-01-15T02:14:54.000Z

What's your computer configuration? I will re-run this experiment and double-check the result.

Answer 6 · 2021-01-15T02:19:28.000Z

Thanks for your reply,my computer configuration is ubuntu 16.04 16G RAM NVIDIA GeForce 1070 pytorch 1.4 cuda 10.1 .thanks.

Answer 7 · 2021-03-02T01:32:11.000Z

please help me on upon questions,thanks

Answer 8 · 2021-03-24T06:43:21.000Z

I have re-run the experiment of r2d2, the results are the same. Did you run this experiment following the README.md in 3rd_party/r2d2 directory?

Answer 9 · 2021-03-24T06:52:12.000Z

The features are extracted from the resized images. The resize operation is done as in line 26 of run_all_evaluations.sh. Please check it.