MeshAnything – Converts 3D representations into efficient 3D meshes
buaacyw.github.io
buaacyw.github.io
Weird custom non-commercial license unfortunately. Notes from the GitHub readme:
> It takes about 7GB and 30s to generate a mesh on an A6000 GPU
> trained on meshes with fewer than 800 faces and cannot generate meshes with more than 800 faces
Its an area where things can be improved a lot imho - I did some work a while back fitting flat planes to pointclouds, and ended up with mesh model anything from 40x to 100x smaller data than the ptcloud dataset. see quato.xyz for samples where you can compare the cloud, the mesh produced.. and view the 3D model in recent browsers.
My approach had some similarity to gaussian splats... but using only planar regions .. great for buildings made of flat slabs, less so for smooth curves and foliage.
Applying their MeshAnything algo to fine meshes from photogrammetry scans of buildings would be of great benefit - probably getting those meshes down to a size where they can be shared as 3D webgl/threejs pages.
Even deciding on triangle points to efficiently tesselate / cover a planar region with holes etc, is basically a knapsack problem, which heuristics, monte-carlo and ML can improve upon.
Basically does a lot of matmulls and sampling to find good good planar fits...
Most of these sample datasets took an hour to compute, using 12 threads on a midrange CPU... should be doable in seconds if ported to a highend GPU.
There are many uses for this tech, particularly in less techy crowds that still make significant use of traditional photogrammetry.
Your solution, if local, could give a significant advantage over other products such as Polycam. Again, if local, you could allow for much bigger scans (wink for those doing architecture, particularly those in the restoration field). Anyhow, hope you get that funding!
You just dump the pointcloud to a hierarchical octree and at viewer end just download the nodes in your frusttum and voila.
There are other approaches but this wins hands down on usability/simplicity.
Also added code to import e57 cube panoramas ..
Still, I think if one can use ML to simplify a pointcloud or fine mesh .. then the data is much smaller and cleaner, easier to import to existing CAD tools etc.
But surfels seems like a clever idea, pre-dating view dependent NeRFs.
Still triangles rather than polygons, but we are getting closer.
The end goal should be:
1) Polygons, mostly 4 sided, rather than triangles.
2) Edge smoothness/creases to separate hard coders from soft corners. (Which when combined with polygons enables SubD support: https://graphics.pixar.com/opensubdiv/docs/subdivision_surfa...)
3) UV for textures that are aligned with the natural flow of textures on those components.
4) Repeating textures (although sometimes not) that work with the UVs and combine to create PBR textures. (Getting closer all the time: https://gvecchio.com/stablematerials/)
After the above works, I think people should move on to inferring proper CAD models from an image. Basically infer all the constraints and the various construction steps.
Besides, merging them unintelligently still doesn’t necessarily form clean edge loops, which are required for good model topology.
So the scale shown in this paper feels like toys! Not undermining the effort at all. We need to start somewhere anyway.
For the same reason, I feel puzzled looking at Industrial scenes in Video Games. They are like 3 order of magnitude simplified compared to a real plant.
Because they are games, not oil refinery simulators. They are typically intending to only convey a general sense of “industrial environment” and nothing more.
Do your models of oil refineries include the correct grass and other plant species growing in cracks in the pavement?
That's an excellent point. I do feel compelled to mention the exception of oil refinery simulator games. Maxis (of SimCity, The Sims fame) made SimRefinery way back when.
Most people don't care as they don't have deep knowledge of how a castle or a power plant really functions, you only notice oversimplifications in media in the field you work in.
It's also very likely the designers and artists didn't have time to do much research, and the whole thing is based off a Pinterest reference board.
ThEy NeVeR AniMaTe The FlAps!!
It's like you'd have an animated motorcycle scene and they don't turn the handlebars or make the bike lean when going around a corner. Like, the graphics are _soooo_ good but then they make the danged plane turn and immersion breaks (for me).
https://www.youtube.com/watch?v=SBHPSmsrEC0
I have never seen flight discussed in such detail, especially by someone who is an artist, rather than a pilot or an engineer.
- sensors used
- postprocessing
- registration algorithm(s)
Are all things that would interest me greatly :)
Really? You don't know why video games don't have 80 billion points and you don't know why a tool made to simplify meshes into video game objects isn't using your 80 billion point lidar scan?
For starters, these are meshes and you're talking about points. If anyone is meshing those points and they have any sense, they are working with "toy" sized chunks too so they avoid doing nearest neighbor calculations on terabytes of data.
That last statement is worded in such a weird way, lol. Funny Chinese->English transliteration.
"The FBI has issued a warning for potential criminal activity resulting from the automatic generation of low-poly models. The public is advised to minimize outdoors exposure and report any suspicious activity."
https://github.com/wjakob/instant-meshes
Cheers =)
Or do you mean something like "contains implied ngons because of the way coplanar tri faces are arranged"?
Further more you would only allow quads to meet in 3, 4, or 5 edges per vertex. The 4 edges per vertex is the "normal" case that most of your mesh should have, it causes a regular grid of parabolic (euclidian) geometry with neutral curvature. Then patches of these meet in vertices with 3 edges to make it elliptic geometry with positive curvature or 5 edges to make it hyperbolic geometry with negative curvature.
You can ignore all of these and just randomly connect nearest neighbors to form triangles. But, then you still have only geometry, no useful topology, so not any better than a point cloud. A good topology is necessary for texturing, skinning, animation etc.
> In 5 of my 10 test there were n-gons. There are always starfishes with 5 or more connected verts.
Ok. So you are basing your definition on the number of edges that meet at a vertex. My understanding was that the important metric was "number of edges on a given face"
The triangle topologies in this paper made don’t follow the logical loops that an artist would work as. Generally it’s rare an artist would work directly in triangles, versus quads. But that aside, you’d place the loops in more logical places along the surface.
The face and toilet really stand out to me as examples of meshes that look really off.
Anyway, I think this is a good attempt at a reasonable topology generation, but the tag line is a miss.
Beside good topology is dependent on the use case, it’s very different if you are doing animation, a 3D print, a game or just a render.
Low-poly re-mesh tools have been around for ages (some better than others), but there are good reasons pro's still do this step manually all the time. Primarily "good" is based on _where_ the quads, loops, and unavoidable n-gons end up in the model (or stuff ends up looking retro 90's.)
There is also the complex legal side of algorithms not being able to create copyrightable works in some jurisdictions. Talk with your IP lawyer, this area gets messy fast when something famous or trademarked is involved.
Cheers, =3
Imagine recreating part of real life city, creating a digital twin, for scientific purposes (testing human behaviour in fire hazards, or simply iterating on better park planning and road design for greater perceived safety). There's a lot to be done, and it's difficult to use procedural building methods if your aim is for people to recognize that area.
I'm making such a thing myself, purely academic, but god I wish I could speed things up.
The problem is when groups start gleaning styles and artwork from 3rd parties to make something in the same style... they cross an ethical line, and a legal one in some situations (even if the original work is completely isolated from the output.)
Thus, while a stochastic parrot may be able to dodge outright plagiarism, it cannot sidestep copyright laws in some Markets.
I'd rather pay folks for royalty free content like Poly Haven offers to the community. =3
I may have not worded things well, I was trying to speak of modern 3D reconstruction methods, such as NeRF or Neuralangelo. I can see good uses for them, as I need to fool the senses reliably (participants will be taken to a VR world... mimicking a real place). But as many things in this field, the reality is that these methods aren't up to snuff. Still, it would be nice to be able to capture reality for non commercial purposes.
As for Polyhaven, haven't donated yet... but I hope to do so soon :)
https://huggingface.co/spaces/Yiwen-ntu/MeshAnything
on the provided sample "hat". I tried with and without checking "Preprocess with marching cubes" and "Random Sample". Both outputs had holes in the output mesh where the original did not.
Am I doing this wrong, or is the algorithm buggy?
The MeshAnything demo is at
Sentiments aside, that's an impressive approach.
https://github.com/buaacyw/MeshAnything/blob/main/LICENSE.tx...