Poor mesh quality - Troubleshooting steps

[Ryan White]

Hi,

I'm having trouble making a decent mesh from a set of images that I imagine "should" work? This is my third or fourth attempt at using Meshroom and it's so promising, but I feel like I'm getting something really wrong somewhere because it never quite produces the results I've seen others able to produce.

I can see that there's multiple similar posts on this group that haven't really had much response. I wonder if there's some way to share a set of images and a graph, and have experts comment on potential solutions (and build up a database/knowledge base of the results?)

At the moment, there's a lot of 'knobs' to tweak, and given the processing time it's difficult to narrow down which ones I should tweak first. I think that's really my main problem.

Anyways, I've attached a screenshot of my problem. There's no mesh at the rear of the boot... Any tips?

Thanks,

Ryan

[Ryan White]

(I should say, the settings in this graph are not any better or worse than any others I came up with, so if there's 'glaringly wrong' stuff, that's probably because I'm just buttonmashing in lieu of knowing what I'm doing. I'd be happy to share the image set, if someone can recommend a convenient way?)

[Ryan White]

I just found this, which might be helpful? - https://github.com/alicevision/meshroom/wiki/Reconstruction-parameters

[David Cattermole]

Hello Ryan,

The most important and often overlooked issue is the photos - everything is derived from the photos. If your photos are not ideal you'll struggle to get good results, no matter the software, or in other words: "Garbage in, garbage out".

From the screenshot it looks to me like your subject (shoe) is too small in frame and there are numerous reflection highlights on the objects that are being tracked (including the table).

Make sure your subject fills the frame. Move the camera closer to the subject (don't just zoom in).

Also, more photos is usually better than less.

To reduce reflections either choose different surfaces, or if you cannot use circular polarizers on lights sources and the camera lens. Other techniques for very shiny surfaces include using talcum powder to "dust" the surface so it includes matte surface details in addition to using polarizers.

I can see you're using CCTags. May I ask why? I'm not sure they will contribute to the solve, but I personally don't have much experience with them.

I'm no expert in Meshroom, so please take my advice with a grain of salt. I cannot give much advice on specific parameters to use - sorry.

Good luck, 

David

--
You received this message because you are subscribed to the Google Groups "AliceVision" group.
To unsubscribe from this group and stop receiving emails from it, send an email to alicevision...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/alicevision/6c9e770f-0fcd-42ae-b878-2dda5e6fc3cdn%40googlegroups.com.

[Ryan White]

Thanks for getting back to me.

I understand for sure garbage in, garbage out. I did try to get what I could this evening in the hopes of getting to grips with the software.

Ultimately, I'd like to use this tool to create a 3D model of the underside of my motorhome, to create drawings for fabrication of a storage unit where a generator used to be. That's why I used CCTags. In that application, I intend to place some CCTags at known distances in the photos, in order to scale the resulting mesh as accurately as possible. One other thing I'd like to consider for that application is using a multi-camera rig. For me, it makes sense that the more things I can control in the setup that's taking the photos, the easier it should be for the software to reconstruct.

Anyways, I'll keep chipping away at the tips in that 'reconstruction parameters' page, and see if anything yields improvements. I'm already onto the first attempt since I found it, and...guided matching (or some other thing I clicked) really slows stuff down.

Thanks again,

Ryan

[Ryan White]

So, I left it running overnight, and it's done the same thing. I set the describer types to dspsift (and cctags), density and quality high, and in feature matching I turned on guided matching.

It seems that the camera positions are still well localized, but for some reason the mesh ends up with holes in it.

I'm trying again with SGM: Nb Neighbour Cameras set to 5 and Refine: Nb neighbour cameras set to 4. I doubt it will work.

One thing I noticed, there seems to be quite a bit of 'noise' in the StructureFromMotion visualization. There's points over the top of the table where there was only empty space in reality. Could that be related?

Do I just need better lighting and more pictures? For my eventual use case, that may be difficult. Would a ring light work for diffuse surfaces, rusty or matte painted metal?

Thanks,

Ryan

[Ryan White]

I managed to reduce the 'noisy points above the table' by increasing 'Min Observation for Triangulation' at the SfM node from 2 to 4. It still reconstructed all views. Re-meshing now.

R