Johan Edstedt comments

Results 235 comments of


                                            Johan Edstedt

Training set and test set setting?

Yes, this seems reasonable. When I ran the original test run I think I got lucky with the auc5. There might also exist some tiny differences due to updates to...

Training set and test set setting?

Hi @stilcrad numbers are referenced from the original paper. Typically this is ok. Agree that it’s unfortunate that pmatch code is unavailable. Good luck with the paper 😀

Training set and test set setting?

I thought you got > 0.622, 0.764, and 0.862 5% lower shouldnt be possible, no.

Training set and test set setting?

Yes, this seems reasonable. Its not 5% difference, its 0.4% diff.

Training set and test set setting?

@nfyfamr I dont have my machine in front of me atm so can't rerun the tests myself, but here's a checklist for what should be enabled to reproduce results. A)...

Training set and test set setting?

Are you using our train split or all scene infos? All scenes in our train split should have undistorted sfm scenes. Your undistorted sfm com3s from running undistort on something...

Training set and test set setting?

> @Parskatt Thanks for the advice. I checked them and used clean conda environment. But the result was the same before. Could you let me know the commit id you...

Training set and test set setting?

Oh, then some bug must have been introduced...

Training set and test set setting?

@nfyfamr Can you verify whether you ran mega_1500 or mega1500_poselib in your eval on the main branch? I saw that I accidentally pushed mega1500_poselib as default. There might be some...

Training set and test set setting?

I should probably add some CI or similar to force test to be repros before pushing to main... Lesson learned maybe.