Pytorch version training

[prbonn]

Pytorch version training

[BIT-MJY]

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

[BIT-MJY]

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]

@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]

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]

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?

[surajiitd]

Thank you so much @BIT-MJY