Why can't I reproduce the results from the paper?

Open Inverse-function opened this issue 1 year ago • 3 comments

I have followed your advice for training, but why are my results three percentage points lower than those reported in the paper, and the results are more similar to the LOFTR model? 561b6751-c2cb-4a97-bfc9-15620db29e77

Dec 27 '24 02:12 Inverse-function

Thanks for you question! The AUC in the checkpoint name is based on testing at the same resolution as training (832) for MegaDepth. To achieve results consistent with the paper, you should utilize the scripts located in the scripts/reproduce_test directory, which test at 1152 resolution for MegaDepth.

Dec 27 '24 06:12 wyf2020

Have you followed the author's advice? Can Auc20 achieve an accuracy of 83%?

Mar 14 '25 02:03 xml-0506

Was this training result trained on the MegaDepth dataset?

Aug 14 '25 07:08 zhouh02