Refine mesh - no updates

[cbnewham]

I've been trying the "Refine Mesh" feature for the first time. It's been going for several hours, not using very much CPU, GPU or main memory (although GPU memory is being used). The console just says "Refining model...".

The total completion bar looks to be at about 5% but remains unchanged since the process did the initial setting up. There is no estimate given for the length of time the process may take - that's just blank.

Does the console ever update at all? There seems to be no debug output for this feature, other than at the very start.

I'm debating whether to stop the process or just let it run - but it could be in an infinite loop for all I know. :-(


Question: Is the process mainly tied to the number of photos, or to the number of faces the mesh has?

[cbnewham]

OK, I've now made some experiments using a small example.

The refine process takes a very long time.

The processing time is mostly bound to the number of photos. The number of faces in the original mesh has very little to no impact on the time.

I can see how useful this is for reproducing fine detail that is removed after performing a smooth operation on the entire mesh. It's a shame it takes so long to process for large image sets.

[Alexey Pasumansky]

Hello cbnewham,

The duration of the refine mesh operation is very tied to the number of images and image resolution (in combination to the quality parameter used), as it actually works based on the image data.

Have you tried to compare the results of your common procedure: build mesh + refine mesh to the results of the mesh generation based on the depth maps in similar quality? Are you getting considerably higher level of detail for your common workflow?

[cbnewham]

Hi Alexey,

My brief experiments indicate the following:

- the overall quality and detail of the mesh is dependent on the initial mesh construction quality (ultra, high, medium, etc).
- details in the mesh can be retrieved by using refine mesh with an ultra or high setting, but course features and overall shape are not affected as much.

It would appear that a good strategy might be to:

1. build the mesh in ultra or high
2. decimate and smooth the mesh to remove noise
3. refine the mesh using ultra or high to bring back details
4. repeat 2 and 3 as needed

I need to perform some more experiments, although I think the refine mesh is outside my capabilities for many of my larger projects - it appears building the mesh is roughly proportional to the number of photos for any given example, while the refine process is not (maybe dependent on the video memory available to the GPU or maybe something else?)

For a 60 camera example I get an ultra mesh build time of 1100 seconds and a refine (10 iterations) of 750 seconds. Yet a big project with over 1000 cameras build the mesh in a few hours while the refine was still going, even after half a day.