static points remove too much

[zjulabwjt]

I use custom driving dataset remove dynamic points, use assert folder config, but I found points around ego car and on the ground is removed. Can you give me some advice to solved this problem?Thank you!

[Kin-Zhang]

It's hard to tell without taking a look at your data. If it's possible could you provide more visualization or data directly?

[zjulabwjt]

It's hard to tell without taking a look at your data. If it's possible could you provide more visualization or data directly?

@Kin-Zhang thanks for your quickly reply!

First image is pcd after remove dynamic points, and second image is origin pcd, third image is removed dynamic pcd. The ground points around ego car is removed too much. I used muti lidars pointcloud and had set different viewpoint in trajectory.Can you give me some advice?Thank you!

[Kin-Zhang]

If that possible to share the data with me and let me know what's your config on it so I can reproduce your results?

[zjulabwjt]

If that possible to share the data with me and let me know what's your config on it so I can reproduce your results?

@Kin-Zhang Thank you for your quick reply! However, it seems inconvenient for me to provide data. May I ask you a few related questions?

1.In my experiments in an underground parking scene, I noticed that a large number of ground points are mistakenly removed on ramps and turns. Is this caused by incorrect parameter settings?

2.Can DUFOMAP handle point cloud results from a single frame?

3.Regarding the issue of ground points being mistakenly removed in the my data, I am currently segmenting the ground and then performing dynamic object removal on non-ground points to avoid the mistake. However, it seems that points near the ground cannot be removed. Is this phenomenon reasonable?

[Kin-Zhang]

1.In my experiments in an underground parking scene, I noticed that a large number of ground points are mistakenly removed on ramps and turns. Is this caused by incorrect parameter settings?

Sorry to say this, but it's really hard to tell if I don't have your data. Especially I don't know if you process data correctly.... As you can see most of issues on their custom data have some mistakes in processing at the end. For DUFOMap I tested on multiple floors (7 floors) in HKUST Academic Hall, and the result was still good enough. So I don't think it's the algorithm problem but may data process.

2.Can DUFOMAP handle point cloud results from a single frame?

Yes, we have kind of single frame result in Python API even. Check this work and Python script.

3.Regarding the issue of ground points being mistakenly removed in the my data, I am currently segmenting the ground and then performing dynamic object removal on non-ground points to avoid the mistake. However, it seems that points near the ground cannot be removed. Is this phenomenon reasonable?

I would strongly recommend octomap w fg: https://github.com/Kin-Zhang/octomap. It's our first try on ray-casting with noise filter and ground segmentation involved. However, I didn't engineer optimized runtime. So the runtime is quite slow compared to DUFOMap. And you can reference the link I provided and see how we do non-ground points but still use ground points to do ray-casting.