Open laurenzheidrich opened 3 months ago
Translating my point clouds down by 1.5-2.5m tremendeously changed the reconstruction result to the better. I am closing this issue :)
Translating my point clouds down by 1.5-2.5m tremendeously changed the reconstruction result to the better. I am closing this issue :)
I encountered a similar issue. Could you please explain it in more detail? Is it about lowering the height of the LiDAR?
Yeah so when I looked at my scans in comparison to Kitti, I saw that the ground points are around 1.5m higher than the ground points of Kitti. So as a last straw to grasp, I just translated all my scans down by 1.5m along the z-axis so that the ground points of my scans and kitti scans are somewhat aligned.
after doing that, VDB started to work perfectly. I don’t know why tbh, not sure if there is some parameter set in their code that detects the ground at a certain pre-defined height or something. Would be great if the authors could say anything about this.
Yeah so when I looked at my scans in comparison to Kitti, I saw that the ground points are around 1.5m higher than the ground points of Kitti. So as a last straw to grasp, I just translated all my scans down by 1.5m along the z-axis so that the ground points of my scans and kitti scans are somewhat aligned.
after doing that, VDB started to work perfectly. I don’t know why tbh, not sure if there is some parameter set in their code that detects the ground at a certain pre-defined height or something. Would be great if the authors could say anything about this.
OK,Thank you very much.
Hi & thanks for your great repository!
Using Kitti Dataset I get very good results. Now I am trying to run your pipeline on a custom dataset. While the fine detail of cars is astonishing, the reconstruction of large structures such as the street is quite bad. I have tried different parametrizations but the problem stays the same.
Here are two screenshots of the resulting reconstructed mesh
And here is a screenshot of the the merged point cloud:
As you can see the merged pointcloud (consisting of 50 scans) is very dense and provides great detail, but the mesh is full of holes and missing parts. I have used the standard Kitti parametrization you provided, but played around with the voxel_size, sdf_trunc, min_weight, fill_holes etc. in the range of [x0.1, x10] and the results were always almost the same.
Can you point me in any direction of what I could change in the parametrization / setting etc.?