How to find best parameter values

[capCafu]

Dear Qingwen Zhang,

I have been using DUFOMap as my main moving object removal algorithm. To run it properly, I converted my data into .pcd format with viewpoint information, and the pipeline works correctly.

However, when using the default parameters from the paper, some ghost objects still remain in the output. I attempted to fine-tune the parameters such as resolution, ds, and dp, but in some cases the algorithm over-removes points, which removes valid parts of the scene. So far, I have not been able to find a stable configuration that works well for my scans.

It would be very helpful if you could share any guidelines or recommendations for selecting appropriate parameter values, or how you typically determine the best configuration for different datasets.

You can access a sample of my data here:
https://drive.google.com/drive/folders/1v_eqZas5kDCPlC3c9oohOg_RLoKfzkMY?usp=drive_link

I have also attached an example result below for reference.

Best,
Sagi

[Kin-Zhang]

great, thanks for all these information and data. I will have a look during these days.

[Kin-Zhang]

Hi, Sagi

Here is the best I think I can try, after checking your data: (left your output; right mine)

The parameter I tried (maybe there is a better one since I made it quite aggressive):

I20260306 18:29:57.156403 140469053200192 dufomap.cpp:52] Here is your dufomap config: 
I20260306 18:29:57.156534 140469053200192 dufomap.cpp:53] Set resolution: 0.2, d_s: 0.2, d_p: 1, hit_extension: 1, ray_passthrough_hits: 1
I20260306 18:29:57.156551 140469053200192 dufomap.cpp:54] Set num_threads: 24

python init:

    mydufo = dufomap(0.2, 0.2, 1, ray_passthrough_hits=True) # resolution, d_s, d_p same with paper.

The remaining dynamic points are tricky because those spaces were only observed once, leading to a poor freespace map in dufomap. To improve results, ensure the sensors observe important areas several times. Understanding the underlying dynamic detection logic will help you collect better data for easier cleaning in the future.

Let me know if you have further questions.

[capCafu]

Thank you for your response. It seems that using more aggressive filtering may remove some small objects that I would like to preserve. For now, it appears that adjusting the scanning strategy—such as performing a denser scan—might help reduce the problem.

I also have one more question: do the LiDAR specifications affect the performance of the overall algorithm?

[Kin-Zhang]

Actually, I do not think the LiDAR specifications directly affect performance. They mainly affect the free space that can be reconstructed. As shown in the paper, we used the same parameters across different LiDAR setups and scan patterns at similar scene scales. Since DUFOMap treats the input simply as points for ray-casting, it is not tied to a specific sensor configuration. In my opinion, the key factor is whether enough free space is observed.

[capCafu]

Thank you for the clarification.
As I have understood it depends on how well free space can be reconstructed and key factor is having enough observations of the scene.

Thanks again for the helpful explanation.