Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Pytorch version training #35

Closed
surajiitd opened this issue Oct 27, 2022 · 8 comments
Closed

Pytorch version training #35

surajiitd opened this issue Oct 27, 2022 · 8 comments

Comments

@surajiitd
Copy link

surajiitd commented Oct 27, 2022

Thank you for sharing your pytorch code. @Chen-Xieyuanli @laebe
In your recommended Data Structure how to generate loop_gt_seq00_0.3overlap_inactive.npz file for training?
I have used the tensorflow branch to generate the preprocessed data, but no such file is generated.

@BIT-MJY
Copy link
Collaborator

BIT-MJY commented Oct 27, 2022

Hi @surajiitd Thanks for using our code!

loop_gt_seq00_0.3overlap_inactive.npz contains N lists for N laser scans of sequence 00. In each list n, there are indices of laser scans that have overlap > 0.3 with respect to the n th laser scan (exclude near 100 scans). As to the calculation of overlap, you can refer to OverlapNet Demo4.

BTW, you can find our newer works OverlapTransformer and SeqOT to implement faster retrieval. We have released the code to generate distance-based loop git files for NCLT dataset here. The data structures are quite similar. The only difference is that each list in git files for NCLT contains the loop indices in database.

@surajiitd
Copy link
Author

surajiitd commented Oct 29, 2022

@BIT-MJY Thank you so much for your instant reply!
But I already ran Demo4 of OverlapNet and tried again. But I can't generate any loop_gt_seq00_0.3overlap_inactive.npz file. No such file is generated that has such info (one list for each laser scan). Please help on how to get this file!
Of course, I'd like to see your more recent work OverlapTransformer and planning to use it. But that is also referring to Overlapnet for generating gt_file(loop_gt_seq00_0.3overlap_inactive.npz) file.

@BIT-MJY
Copy link
Collaborator

BIT-MJY commented Oct 30, 2022

We have uploaded our newest version of this related script to generate ground truth files in the OT repo. You can find it here.

Please pull our new code, modify the paths in this file (tools/utils/gen_gt_data.py), and run

cd tools/utils/
python gen_gt_data.py

Note that the poses file is from semantic kitti, which is refined by SLAM.

@surajiitd
Copy link
Author

surajiitd commented Oct 31, 2022

Hi @BIT-MJY

Note that the poses file is from semantic kitti, which is refined by SLAM.

In OverlapTransformer you have used semanticKitti poses, but in OverlapNet you have just used kitti poses right?

@BIT-MJY
Copy link
Collaborator

BIT-MJY commented Nov 1, 2022

I suppose so, but we recommend using semanticKitti poses in your work which are more accurate.

@surajiitd
Copy link
Author

Thank you so much @BIT-MJY

@LY0406
Copy link

LY0406 commented Jul 4, 2024

您好,我正在学习您的OverlapTransformer这篇论文,您在论文中给出的新数据集Haomo数据集似乎仅仅给出了激光序列和位姿文件而没有提供训练OT所需要的训练集,如果我想要得到训练集应该怎么做呢。

@laebe
Copy link
Member

laebe commented Jul 4, 2024

LY0406, would be nice to write in english so that everybody (include me) could read it !

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

4 participants