Flipped triangles when simplifying mesh

[andreiKDI]

trafficstars

Hi! First of all, thank you for working on this great tool, much appreciated! 👍

I'm trying to simplify some terrain tiles which are basically just grids with a bit of elevation. The problem is that sometimes, the simplification process generates weird meshes, with some triangles flipped, others overlapping and some very thin. You can see all three of them in the "after.obj" attached. Here's the code I use for simplifying the mesh, although the same happens without the index and vertex remapping, just that the buffers will be bigger. Also, I tried tweaking the threshold and targetError, with similar results.

    // Simplify mesh
    float threshold = 0.2f;
    size_t targetIndexCount = size_t(indexes.size() * threshold);
    float targetError = 1e-2f;

    indexes.resize(meshopt_simplify(indexes.data(), indexes.data(), indexes.size(), &(positions[0].x), positions.size(), sizeof(glm::vec3), targetIndexCount, targetError, nullptr));

    std::vector<unsigned int> vertexRemap(positions.size());
    size_t newVertSize = meshopt_optimizeVertexFetchRemap(vertexRemap.data(), indexes.data(), indexes.size(), vertexRemap.size());
    vertexRemap.resize(newVertSize);

    meshopt_remapIndexBuffer(indexes.data(), indexes.data(), indexes.size(), vertexRemap.data());
    meshopt_remapVertexBuffer(positions.data(), positions.data(), positions.size(), sizeof(glm::vec3), vertexRemap.data());
    meshopt_remapVertexBuffer(texCoords.data(), texCoords.data(), texCoords.size(), sizeof(glm::vec2), vertexRemap.data());
    positions.resize(newVertSize);
    texCoords.resize(newVertSize);

Here you can find the before and after meshes with the fewest grid cells I could reproduce the problem with and without normals, uvs etc: Desktop.zip

I briefly tried to debug the problem myself, but it wasn't obvious to me what happens there. Please let me know if you need more details.

[zeux]

Sorry it took so long, I just resumed working on various simplification backlog items, this being one of them. Unfortunately I'm now having trouble reproducing the results :(

Could you clarify:

• which version of meshoptimizer are you using
• which compiler you're using to build meshoptimizer; in particular whether you're using any special floating point optimization settings such as -ffast-math//fp:fast
• whether you can reproduce the problem in gltfpack by running gltfpack -i after.obj -o new.glb -si X, X being the simplification factor, with your version / compiler?

Sometimes the .obj preprocessing in gltfpack obscures the bugs but it doesn't seem to be the problem here, and I tried running gltfpack in a special mode where the index/vertex buffers from the input are preserved and still can't reproduce this.

[kristoffer-dyrkorn]

It seems I am able to reproduce this bug (ie overlapping triangles) using gltfpack 0.17 on macos (binary downloaded from here: https://github.com/zeux/meshoptimizer/releases/tag/v0.17) and also the latest built npm-based gltfpack (the artifact from here: https://github.com/zeux/meshoptimizer/actions/runs/2531821310, ie https://github.com/zeux/meshoptimizer/suites/7014049981/artifacts/275430512).

Command line (and output) was: ./gltfpack -v -noq -si 0.2 -i test.obj -o test.glb input: 0 nodes, 0 meshes (1 primitives), 0 materials, 0 skins, 0 animations input: 1 mesh primitives (4802 triangles, 14406 vertices); 1 draw calls (1 instances, 4802 triangles) output: 1 mesh primitives (959 triangles, 509 vertices); 1 draw calls (1 instances, 959 triangles) output: 1 nodes, 1 meshes (1 primitives), 0 materials output: JSON 552 bytes, buffers 11864 bytes output: buffers: vertex 6108 bytes, index 5754 bytes, skin 0 bytes, time 0 bytes, keyframe 0 bytes, instance 0 bytes, image 0 bytes

Please see attached file for screenshots and obj file used as input: files.zip

[happyfrank]

I tested a small model and it still seems to have this problem -.- version: version 0.18 void simplify(const Mesh& mesh, float threshold = 0.2f) test_mesh

[zeux]

@happyfrank would you mind attaching the example to this issue as a .zip file? The link is not reachable.

[happyfrank]

@zeux Certainly, I can confirm that the link is now working and you can access the example. Please try accessing it again and let me know if you encounter any issues. test_mesh

[zeux]

@happyfrank I can download this now, but I'm having trouble locating the geometry with the overlap in Blender in the simplified scene, and unfortunately MeshLab (which I think the screenshot is from) crashes when trying to open it :-/ If you could upload a separate .obj file with just the geometry with the overlap shown in the screenshot, that would help - I'd normally be able to cut it out myself but I can't quite understand what portions of the screenshot correspond to what objects in the original scene as well :) as it's very chaotic.

[happyfrank]

@zeux I apologize for the confusion earlier. I have created a new .obj file with just the overlapping geometry shown in the screenshot, and you can download it from here: test_mesh2 I also simplified the model to make it easier to work with. Let me know if you have any further questions or concerns. version: version 0.18 void simplify(const Mesh& mesh, float threshold = 0.2f)

[zeux]

Perfect, thanks! I can reproduce this on that file now. Reuploading the file so that it is attached to the issue: mesh_flip.zip