[request] (really) dense point cloud

[andrea-bistacchi]

Hello, a lot of people working in photogrammetry are more interested in a dense point cloud than in a textured surface. However Meshroom yields very poor dense clouds, that are generally between one and two order of magnitude less dense than those obtained with Agisoft or CaptureReality, and also VisualSFM point clouds are one order of magnitude denser.

We suspect that this is mainly a problem of setting the parameters right in the MVS step, and then exporting the full-resolution point cloud instead of a filtered one. Many issues have been opened dealing with this problem (#1462, #1302, #1203, #1132, #888, #885, #803, #690, #382, #266, etc.), but no real solution has been proposed.

Please help us, otherwise Meshroom will be useless for us (and for all people interested in point clouds), and this would really be a pity since it is a great project, really!

[natowi]

You can get a really dense point cloud from the Meshing node as described here: https://github.com/alicevision/meshroom/wiki/Export-Pointcloud

There are also a few settings in other nodes to increase the point density. The pointcloud from the StructureFromMotion node is always less dense than the one from the Meshing node using the default pipeline.

[andrea-bistacchi]

Hello, thanks for answering but this is what we are exporting, and it is not dense when compared to the output by other codes.

If you like we can send some example, but I guess we cannot attach them here (we need some GB).

I guess Meshroom works in principle as the other codes, so we need some advice on the "settings in other nodes to increase the point density".

Thanks very much!

[fabiencastan]

Yes, it would be great if you could share some concrete examples with the number of points you get in the different solutions, so we can reproduce it and provide precise recommendations.

My email to send the datasets: fabien.castan@technicolor.com (you can use fromsmash.com or wetransfer.com).

[andrea-bistacchi]

Hi, one of my students will send all details tomorrow. We are really struggling with all parameters and we cannot find a solution.

If you like we can send the whole projects with photos etc., but in this case we need an email address.

Thanks very much!!!

[stefano-casiraghi]

Hi, I'm the student who works with Andrea. We are mainly interested in the 3D point cloud of natural outcrops. Reading through the user's guide I found a few parameters to work with, for instance in the Feature Extraction node I set the Describer Quality and the Describer Density to High,

and in the Meshing node, I maximize the Max Input Points, Max Points, and Max points per voxel options.

The resulting point cloud is characterized by a very poor point density (490,000 points), and the outcrop surface is noisier if compared to a point cloud coming from a standard Agisoft Photoscan workflow (18,000,000 points).

Meshroom:

Agisoft (Visualized in Cloud Compare):

Any advice will be highly appreciated,

Thanks very much!

[natowi]

I compared the results of Agisoft and Meshroom. Yes, Agisoft has a really Dense Point Cloud by default, but Meshroom achieves a Mesh with a similar and sometimes even better quality:

From my experience, Meshroom focuses on robust reconstructions and discards many points that are not robust enough by default: Areas that were visible in the sparse StructureFromMotion point cloud and DepthMaps don´t make it into the final reconstruction.

From the Meshing log it looks like we internally have a point cloud density similar to Agisoft, but it is filtered down later:

By tweaking the parameters I was able to increase the point cloud density for three images from the monstree dataset from 32K points (default) to 450K points - but the Agisoft point cloud is far denser still (4.8M points)

Here are the parameters I adjusted: Meshing: max input points, max points, max points per voxel - to maximum value and Add Landmarks To The Dense Point Cloud and Colorize Output enabled, Densify enabled, Densify front/back 5 - by adjusting the helper point settings it is possible to recover areas that would otherwise be dismissed like in the images above DepthMapFilter: relax the Min Consistent Cameras and Min Consistent Cameras Bad Similarity to 2 and 3 respectively (completely removing this node can increase the point cloud density with the downside of having noise) DepthMap: Downscale 1 Enable additional describers and adjust the quality and density (depends on your dataset)

If you just need a point cloud (and not specifically the original Meshroom Pointcloud), you could use Meshlab or Cloud Compare or trimesh to convert the final Meshroom Mesh.

[andrea-bistacchi]

Hello, and thanks!

I totally agree on mesh quality. In fact quite often in Agisoft you need to clean the dense point cloud before meshing.

However for some application we do not us the mesh but the point cloud directly, and this can be very dense and still accurate if the photo dataset is very redundant (i.e. thousands of photo, with 6 to 10 photos looking at each portion of the model).

We'll try to work on the parameters you highlighted.

Thanks!

[fabiencastan]

DepthMap: Downscale=1 is the main one. Meshing: PixSizeMarginInitCoef=1, PixSizeMarginFinalCoef=1, MinStep=1,