How to remove noise points?

[NUCIE315]

When training one's own background free dataset, there is noise in the output 2D GS points。And how can I remove them?

[hbb1]

I don't get you point. From my side, it works for object without background. Here is the nerf_synthetic chair for an example. The left is the learned point cloud while the right is the extracted mesh. Both of them are free of floaters. Maybe you case is different.

[NUCIE315]

Thank you for you answer. Below first is my image, it only includes the central object, but the learned point cloud may have some noisy points, as mentioned earlier and I don't know why? When I use a small depth_trunc(3 or below) to extract mesh, I only can get the mesh of noisy points. When I use a big depth_trunc(8 or higher) to extract mesh, I can get the excellent mesh but with noisy points, as shown in the second image. And I use the 3DGS, the learned points don't include nosiy points, as shown in the third image. Is it possible that some parameters were not adjusted well during training? Do you have any suggested solutions or parameters for the above situation？ Thank you!

[hbb1]

You can try to adjust the star iter of the regularization, may be 1500 or larger. And, you can post-processed the mesh to extract one cluster.

python render.py -m <path to pre-trained model> -s <path to COLMAP dataset> --skip_train --skip_test  --num_cluster 1

[NUCIE315]

Thank you very much, I'll give it a try later.

------------------ 原始邮件 ------------------ 发件人: @.>; 发送时间: 2024年5月15日(星期三) 下午4:55 收件人: @.>; 抄送: @.>; @.>; 主题: Re: [hbb1/2d-gaussian-splatting] How to remove noise points? (Issue #31)

You can try to adjust the star iter of the regularization, may be 1500 or larger. And, you can post-processed the mesh to extract one cluster. python render.py -m <path to pre-trained model> -s <path to COLMAP dataset> --skip_train --skip_test --num_cluster 1
— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.***>

[oUp2Uo]

I guess if there was a program like SuperSplat (which is used to edit the generated 3DGS .ply point cloud file) to process 2DGS .ply point cloud file before export mesh, will make the result with less noisy points.