3D scanning is becoming more and more ubiquitous. Robotic automation, self-driving cars and multitude of other industrial, medical and scientific applications require advanced computer vision to deliver the levels of automation customers expect these days. The great rise of AI gave another boost to computer vision and the need to utilize 3D data to make machines smarter. Not only are tasks at hand becoming more complex, but the size of data grows exponentially.

There is a multitude of general purpose libraries which deal with 3D data. Some stem from popular CAD packages, some are open source. The commercial ones tend to be quite expensive while open source are free though tend to be limited in functionality provided. Also those libraries value generality above other features to allow maximum number of applications, but with the growing amounts of 3D data, performance is critical as it never has been. Some of it can be addressed by using the scale of a commercial cloud, last generation CPU or GPU but underlying complexity of data representation makes it very hard and laborsome.

The goal which we set when designing MeshLib was to value simplicity and performance while providing a wide gamut of useful computational algorithms. The library also supports the most important data structures today’s sensors can produce - pointcloud, mesh, volume and more. For example, mesh is represented by half-edge data structure and cannot be made non-manifold. Some applications may require non-manifoldness, but most practical scans can be represented as manifoldness meshes without an issue.

• MeshInspector is a free application based on MeshLib

• MeshInspector online web-browser version (simple email sign-in is required)

• MeshInspector web-site

• Comparison of MeshLib vs VTK library

• MeshInspector YouTube channel

• MeshLib Documentation

• Email us

• Submit an issue

This list is not full and updating each day

• Math primitives and operations support (Vectors 2D, 3D, 4D; Lines; Planes; Bounding Boxes; Matrices; Affine transformations; Quaternions; Histograms; etc.)

• 3D data various representations support: Mesh, Voxel and Point Cloud.
• Data creation
  • Mesh creation by given vertices and triangles,
  • Surface primitives (e.g. tor, cube, sphere, etc).
• Representation conversions
  • Triangulation of a Point Cloud to Mesh,
  • Mesh to Cloud Point conversion,
  • Mesh to Voxel conversion,
  • Voxel To Mesh conversion.
• Deformations
  • Laplassian deformation,
  • Freeform deformation,
  • Relax, mesh smoothing,
  • Position Verts Smoothly, arrangement of vertices in accordance with the gradient along the neighbors.
• Offsets
  • Mesh offset,
  • Mesh partial offset.
• Resolution
  • Breaking a mesh into smaller triangles (increasing the resolution)
  • Mesh decimation (decreasing the number of triangles, decreasing the resolution) with a specified error relative to the old mesh.
• Cutting a contour on a surface. The mesh is modified so that the resulting contour contains new edges cut through triangles at the specified points.
• Splitting
  • Splitting mesh into sub-meshes (components)

• Boolean operations (union, intersection, difference) on 3D objects bounded by meshes. MeshLib has two independent modes:
  1. Boolean ops on meshes via intermediate conversion into voxel representation. This mode requires closed meshes without holes on input and much memory for high accuracy but tolerate for various input mesh degenerations and self-intersections.
  2. Direct Boolean ops on meshes using robust predicates, producing exact accurate result and keeping original mesh elements as is if they are not intersected. This mode supports for open meshes with holes as long as the boundary of the mesh is not intersected.
     • According to our user reports, MeshLib direct Boolean operations are significatly faster even compared to the latest published approaches, e.g. Interactive and Robust Mesh Booleans
• Construction of Convex Hull of a point cloud or a mesh.

• Fixing holes in mesh
  • Holes stitching (removing two holes by stitching their boundaries)
  • Hole filling,
  • Holes fixing metrics
    • Basic set of triangulation metrics,
    • Extended set of triangulation metrics,
    • Custom triangulation metrics.
• Delaunay triangulation optimization, changing triangles without changing vertices, according to Delaunay criterion,
• Tunnels fixing,
• Multiple edges detection
• Degenerate triangles fixing,
• Undercuts fixing, via voxels,
• Surface self-intersections fixing
  • Guaranteed fix via voxels,
  • Fix via Relax (mesh smoothing).

• BVH hierarchies (AABB trees) for meshes and polylines to accelerate all other operations
• Projection
  • Projecting a point onto a mesh - closest point queries
• Intersection
  • Intersection of a ray with a mesh (ray tracing),
  • Intersection of a plane with a mesh, result is a contour,
  • Finding a contour representing intersection of two meshes, meshes remain unchanged,
• Distance
  • Distance from a point to a mesh,
  • Minimal distance between two meshes,
  • Mesh distance map (height map),
  • 2D contour distance map.
• Segmentation
  • Semi-auto voxel segmentation (volumes classification).
  • Semi-auto mesh segmentation by curvature.
• Sampling
  • Mesh sampling. The result is a separate thinned set of vertices, the mesh remains unchanged.
  • Point cloud sampling. The result is a separate thinned set of points, the cloud remains unchanged
    • Uniform cloud sampling,
    • Grid cloud sampling.
• Path
  • Finding a shortest path through the mesh vertices,
  • Finding a geodesic shortest path on the surface, not necessarily through mesh vertices.
• Iterative Closest Points (ICP), two meshes aligning
  • Point to point,
  • Point to plane.

We plan to add computed-tomography reconstruction in MeshLib (already present in MeshInspector). Please write us if you like to see some other features.

Although MeshLib is written in C++, all functions are exposed to Python as well. Python code can be executed both from within a MeshLib-based C++ application (embedded mode) and from python interpreter directly, which imports MeshLib modules.

Moreover, the latest MeshLib released version can be easily installed as a Python 3.8 - 3.12 package using pip install:

• On Windows via

py -3 -m pip install --upgrade pip
py -3 -m pip install --upgrade meshlib

• On Linuxes supporting manylinux_2_31+, including Ubuntu 20+ and Fedora 32+ via

sudo apt install python3-pip
python3 -m pip install --upgrade pip
python3 -m pip install --upgrade meshlib

• On macOS 12.0+ via

pip install --upgrade pip
pip install --upgrade meshlib

See Python Basic Examples here or search for more complex ones on stackoverflow.com.

MeshLib can be build on Windows using either Visual Studio 2019 or Visual Studio 2022, both of which support c++20 language standard.

git clone https://github.com/MeshInspector/MeshLib.git
cd MeshLib
git submodule update --init --recursive
cd ..

Note: following below steps will take about 40Gb of your disk space.

Some third parties are taken from vcpkg, while others (missing in vcpkg) are configured as git submodules.

Windows version of MeshLib is configured to use

• CUDA v11.4 in Visual Studio 2019 and
• CUDA v12.0 in Visual Studio 2022.

Please install CUDA from the official site. If you would like to use another version of CUDA, please modify MRCudaVersion in MeshLib/source/platform.props.

1. Please install vcpkg, and integrate it into Visual Studio (note that vcpkg requires English language pack in Visual Studio, and vcpkg cannot be installed on FAT volumes, only on NTFS):

   git clone https://github.com/Microsoft/vcpkg.git
   cd vcpkg
   git checkout 2024.03.25
   .\bootstrap-vcpkg.bat
   .\vcpkg integrate install (with admin rights)

   More details here: vcpkg.

2. (Optional, but recommended) Install AWS CLI v2. Once installed, reopen PowerShell or CMD. This will allow you to use the vcpkg binary cache from our aws s3, which will speed up the installation process and reduce the required disk space.

3. Execute MeshLib/thirdparty/install.bat having previously installed vcpkg as the current working directory (or simply add vcpkg to PATH environment variable).

4. Open solution file MeshLib/source/MeshLib.sln in Visual Studio. Build it and run.

We regularly check that MeshLib can be built successfully on Ubuntu 20.04 LTS, Ubuntu 22.04 LTS (both x64 and ARM), and Fedora 37.

Install/Build dependencies. Build project. Run Test Application

Install CUDA v12.0 from official site

Use automated installation process. It takes ~40 minutes if no required packages are already installed. This approach is useful for new MR developers

Run the following in terminal:

git clone --recurse-submodules https://github.com/MeshInspector/MeshLib.git
cd MeshLib
./scripts/build_thirdparty.sh  # do not select emscripten in the corresponding question
./scripts/build_source.sh  # do not select emscripten in the corresponding question
# create and install package for Ubuntu
./scripts/distribution.sh
sudo apt install ./distr/meshlib-dev.deb
# create and install package for Fedora
./scripts/distribution_rpm.sh
sudo apt install ./distr/meshlib-dev.rpm

NOTE: ./scripts/install*.sh scripts could be used as well, but apt install is preferable.

NOTE: ./scripts/install_thirdparty.sh script copies MR files directly to /usr/local/lib. Remove this directory manually if exists before apt install deb package.

NOTE: You could specify build type to Debug by export MESHLIB_BUILD_TYPE=Debug. Release is default. Set MESHLIB_KEEP_BUILD=ON to suppress full rebuild

This installation was checked on Ubuntu 22.04 with Emscripten 3.1.48.

Install Emscripten (find more on Emscripten official page)

git clone https://github.com/emscripten-core/emsdk.git
cd emsdk
git pull origin # optional
./emsdk install 3.1.48 # (or another version / latest)
./emsdk activate 3.1.48
source ./emsdk_env.sh

Build

cd ~/MeshLib
./scripts/build_thirdparty.sh # select Emscripten in the corresponding question
./scripts/build_source.sh # select Emscripten in the corresponding question

NOTE: Set MESHLIB_KEEP_BUILD=ON to suppress full rebuild

Run

python3 -m http.server 8000 # note that server should have specific COEP and COOP policies for multithread version
# open in browser 127.0.0.1:8000

You can download dev package and install it with your Linux OS package manager. Examples of integration with cmake can be found in the ./examples directory.

Same as for Linux, but requires brew packages to be installed:

xargs brew install < /Library/Frameworks/MeshLib.framework/Versions/Current/scripts/macos.txt

There are two general options of integrating MeshLib into your project:

1. Submodule
2. Distribution

Common for both options: install thirdparty

You can have MeshLib as submodule in your repository, and inculde all MeshLib's projects to your solution.

NOTE: You should use MeshLib/source/common.props in other projects of your solution.

NOTE: You can customize props by defining CustomMRProps.props in directory above common.props

You can download distribution and integrate it in your projects.

Project settings:

1. C/C++ -> General -> Additional Include Directories add distribution\install\include;
2. Linker -> General -> Additional Library Directories add distribution\install\app\$(Configuration);
3. Linker -> Input -> Additional Dependencies add distribution\install\lib\$(Configuration)\*.lib;
4. Debug: C/C++ -> Preprocessor -> Preprocessor Defenitions add _ITERATOR_DEBUG_LEVEL=0;
5. vcpkg -> Triplet set x64-windows-meshlib

Make sure you copy all dlls from distribution\install\app\$(Configuration); to your $(TargetDir)

NOTE: MeshLib distribution has x64 build only

NOTE: Distribution is build with ITERATOR_DEBUG_LEVEL=0 in debug so you will need to setup this for your projects