Photogrammetry mesh vs. laser scan cloud comparison

[ujvarizs]

Hi,

We are researchers currently working on a project, and we would like to compare a point cloud created with a Leica laser scanner, with a photogrammetry model created by RealityCapture. Our goal is to establish how much the photogrammetry scan deviates from the laser scanned point cloud, and get some good metrics & insight about the differences. We are also fairly new to Cloudcompare, so we have a few questions which we were unable to properly answer based on the tutorials and prior forum posts.

What we have done so far is importing both models, and then computing the normals with preferred orientation. Then we align the laser scanned point cloud with the photogrammetry mesh (first manually, then with the fine registration tool), and we do the segmentation with both the cloud and the mesh selected to make sure that the same relevant part is cut from both models.

Then, we use the cloud-to-mesh feature to compute the differences. However, we get different results based on
- which model we select as „aligned” and as „reference” during registration
- which model we select as „compared” and „reference” during comparison
- if we create a point cloud by sampling the mesh via the „sample points on a mesh” function, and compare the sampled cloud with the laser scanned cloud instead of comparing with a mesh.

It is also our understanding that we should compute unsigned distances in order to get a good overview about the absolute differences between our compared mesh and cloud.

Also, our meshes are unclosed, so after computing the normals with the preferred orientation, the back of the mesh turns completely black. We assume that this means that the blacked-out portion is disregarded during computation.

We would like to know if our workflow is correct, and gain a better understanding about how picking and swapping the „reference” and „aligned/compared” models during registration/comparison affect the results. We are also curious if maybe we should create a point cloud by sampling the mesh and compare that to the laser scanned cloud instead. We are also unsure if our assumptions mentioned above are correct.

We appreciate any help you can provide.

[daniel]

So definitely, you have to chose wisely which entity is the 'reference', and which one is the 'compared/aligned' one. As these operations are not symmetrial (as you are computing the distances from entity A to B, which potentially have different coverage, density, etc.)

Then, I don't think you need to compute normals (they are not used for registration, and they are not used in C2M distances computation, i.e. with a mesh as reference). But to answer your question, no the normals are used even if they appear black, the fact that they appear black just means they are pointing in the opposite direction.

If you do have a mesh (and not a point cloud), then I would advise to use it as reference (as you won't get any issue with point density / sampling). And I personnaly prefer to have signed distances, as you can detect a potential shift / bias (with the average) and still get the min and max distances (from whch you can derive the absolute max). While with unsigned distances, you won't spot a bias, and you will obtain a folded distribution (Normal or not) which is a strange beast, from which the average and std. deviation is not what one generally expects.

[ujvarizs]

Dear Daniel,

Thank you very much, your answer is much appreciated. We have one more question. Looking at the heatmap after the C2M comparison, when we rotate the model, the unclosed side (the inside of the model) is showing different colors ( greater differences ) than the surface of the model. Since the unclosed side/inside of the model is completely irrelevant, we are afraid that this could negatively influence/distort the results of the comparison. Is that the case, and if yes, is there maybe a way around this?

[daniel]

Can you maybe post a snapshot of this? It's not super clear to me...

[ujvarizs]

Dear Daniel,

Thank you for the quick answer! Here is a snapshot of the result of a C2M comparison with the mesh as reference. Looking at the heatmap from the front (left side on the snapshot) seems to show different colors (much fewer red areas) than looking at it from the unclosed side / inside (right side on the snapshot). It appears to us as if the results were a bit different on the unclosed side/inside. Maybe it is just the way the heatmap is displayed.

https://drive.google.com/file/d/1GE8Zkg ... sp=sharing

[daniel]

The colors can change if you have multiple points covering themselves. But it's hard to tell, even with these snapshots... Anyway in this case it means that indeed the overall results are impacted by these multiple layers / internal stuff.