Some questions.

[prbonn]

Some questions.

[Chen-Xieyuanli]

Hey @kissb2,

Thank you a lot for following our work and for your great questions!

1. The semantic information is a kind of extra information from learning-based semantic segmentation which will definitely help the network to distinguish different places better. For a more detailed ablation study, you could find it in our paper Sec IV-E.

2. We didn't make any experiments like that so far. I guess it will still work to predict the overlap between two LiDAR scans. If the scenes are not too different, a loop closing/ICP still may work. If you use different robots, then the mounting height of the LIDAR is important: If the height is too different, then the scans may have no/low (ground-truth) overlap even when you are at the same position and the scene has not been changed at all. Here I assume you use a lidar with 32 or 64 beams that have a small vertical field of view. Therefore an ICP between such scans may also be difficult. If the height is the same (and the lidar is not tilted), it should work.

3. We test our method on KITTI and Ford campus datasets, which are collected from the real world with dynamic objects and it works well. We didn't explicitly design any architecture to remove dynamic objects while trained our light-weight network in an end-to-end fashion. I think the multiple cues from the LiDAR scan will help the network resist the dynamic objects. One could also explicitly use geometric and semantic information to get rid of the dynamic objects like what we did in our SuMa++ to improve the overall odometry and mapping results.

I hope this helps!

[gisbi-kim]

Thank you for the kind explanation!