Dear @Author, thanks for sharing such impressive work! But I've got some doubts:
In the latest minor enhancement revision(https://arxiv.org/pdf/2007.01595.pdf) of your work, it specifically mentioned that searching in the feature space should be limited within an adjustable distance threshold of the odometry position. I assume that is because they try to speed up the searching or the global searching is not very robust.
Besides, in the Fig.10 of your IJRR paper, it shows only 50% of correctly localized poses. It is not even SOA even considering only 65% of query poses are within 50m of the target map. For comparison, scan context achieved >95% true positive in KITTI00.
Recently I am trying to do global localization in NCLT datasets(http://robots.engin.umich.edu/nclt/) borrowing your method, but I still have doubts about the actual performance when it comes to searching in the whole feature space.
Dear @Author, thanks for sharing such impressive work! But I've got some doubts:
In the latest minor enhancement revision(https://arxiv.org/pdf/2007.01595.pdf) of your work, it specifically mentioned that searching in the feature space should be limited within an adjustable distance threshold of the odometry position. I assume that is because they try to speed up the searching or the global searching is not very robust.
Besides, in the Fig.10 of your IJRR paper, it shows only 50% of correctly localized poses. It is not even SOA even considering only 65% of query poses are within 50m of the target map. For comparison, scan context achieved >95% true positive in KITTI00.
Recently I am trying to do global localization in NCLT datasets(http://robots.engin.umich.edu/nclt/) borrowing your method, but I still have doubts about the actual performance when it comes to searching in the whole feature space.
Thanks for your time!