ethz-asl/dynablox

Questions

gisbi-kim opened this issue · 7 comments

gisbi-kim commented

Thanks for the great work and sharing.

I have a couple of questions

  1. outdoor results or videos are available?
  2. any tips to tune parameters for outdoor (i.e., more sparse and fast dynamic objects, even more ego pose errors) datasets?
Schmluk commented

Hi @gisbi-kim

Thanks for reaching out!

  1. Some of the DOALS as well as our own sequences are outdoors, some of them are depicted in the paper and video.
  2. Can you be more specific? We ran the same parametrization outdoors and it worked well in our cases. Also fast moving objects should not be a problem.

Hope this helps!

gisbi-kim commented

I would like to test the dynablox for my own fast moving driving cars (e.g., 30-60km/h) in urban environments. i.e., the moving objects contains mostly cars, as well as the people, as like in DOALS dataset (ps. I tested the DOALS dataset and I love the awesome result on it). The height of the top lidar horizontally mounted on a car is particularly higher than that of the DOALS dataset (e.g., 2-3m). In addition, I am curious about how robust it will be to pose estimation errors (for example, since there may be position errors of several tens of centimeters, especially for high-speed vehicles).

Of course, I don't mean to imply that Dynablox should solve all of these issues; rather, I just enjoy discussing my curiosities casually with the academia :) thanks for the fast response.

Schmluk commented

Thanks for your interest!

We have not evaluated the method in driving scenarios since our main application is robotics, please feel free to test it, we'd be excited to hear how it goes!

Some expectations w.r.t. your questions:

  • The method does not depend on moving object velocity (unless it's very slow), so detecting fast objects should not be a problem.
  • The height and pose of the sensor should also not make a difference as long as the area of interest can be seen and mapped sufficiently.
  • We show in the paper robustness w.r.t. pose estimation errors for high drift rates (up to 3.5 m/s in our experiments). This holds as long as the pose errors are of a drifting (i.e. incremental accumulation of errors) nature. For jumps of 0.5 m every frame I'd expect some wrong detections.

Hope that helps and let us know if you decide to test the method on your data!

  • 👍1
  • ❤️1
gisbi-kim commented

@Schmluk , thx for the very kind tips!

Kin-Zhang commented

I refactored all code and tried on KITTI without ros:

real_detect-2023-05-04_21.34.25.mp4

Maybe need some parameter tunning suggestions, Hahaha.. 😀

  • 👍2
  • ❤️2
muzi2045 commented

looks like the param need to be modifyed in urban environment. LOL +_+. @Kin-Zhang

YZH-bot commented

Thank you for the excellent work.
I have noticed that the mapping performance is some kind suboptimal in highly mobile scenarios. Do you have any suggestions regarding this scene?

dynablox_on_kitti_05.mp4