search

drormoran / Equivariant-SFM

MIT License
87 stars 15 forks source link

about results #4

Closed ajunyo closed 2 years ago

ajunyo commented 2 years ago

it's a nice job,but I found that the results were bad that the point cloud couldn't even produce a normal shape，Do I need to modify any super parameters。and，I did two experiments, using Locus point and not using Locus point，but it seems the same.thanks! 5` (Q(NP4_8_V9%%N~@U59B ？

drormoran commented 2 years ago

Hi,

The current parameters should reproduce the results that appears on our paper. Can you give me some more details on what you are trying to do? Is it a single scene reconstruction or a trained model? Does this problem occurs on the reconstruction of this specific scene or for all your reconstruction?

What do you mean by using Locus point?

ajunyo commented 2 years ago

Thank you for your timely reply. Sorry to disturb you，

i use

python multiple_scenes_learning.py --conf Learning_Proj.conf to train this model,This is almost the result of all scenes。

Locus point i mean use_gt=True

should i change some thing in conf?

thank you!!!

Original

From:"drormoran"< @.*** >;

Date:2022/2/16 15:40

To:"drormoran/Equivariant-SFM"< @.*** >;

CC:"我自己的邮箱"< @. >;"Author"< @. >;

Subject:Re: [drormoran/Equivariant-SFM] about results (Issue #4)

Hi,

The current parameters should reproduce the results that appears on our paper. Can you give me some more details on what you are trying to do? Is it a single scene reconstruction or a trained model? Does this problem occurs on the reconstruction of this specific scene or for all your reconstruction?

What do you mean by using Locus point?

&mdash; Reply to this email directly, view it on GitHub, or unsubscribe. Triage notifications on the go with GitHub Mobile for iOS or Android. You are receiving this because you authored the thread.Message ID: @.***>

drormoran commented 2 years ago

There is an ambiguity in the uncalibrated setup reconstruction. In this setup, you can apply any Homography on the scene and still have a good reconstruction. This is why your plot seems incorrect even when the reprojection error is low. This is why we usually don't look at those plot in the uncalibrated setup.

If you wish to reproduce the reconstruction shown in our paper you need to run your experiments on the calibrated setup, where this ambiguity doesn't exists.

ajunyo commented 2 years ago

Thank you very much for your timely reply. It's really an excellent job

drormoran commented 2 years ago

Thanks :)