opensourceprojects.dev

A broadsheet for software that doesn't ask for your email

OpenMVS fills the Multi-View Stereo gap after Structure-from-Motion
GitHub RepoViews10

Project Description

View on GitHub

OpenMVS: The Missing Link Between Camera Poses and Textured 3D Models

You've got a folder of photos, and you want a 3D model of the scene they capture. If you've been down this road, you probably know about Structure-from-Motion tools like OpenMVG—they'll recover camera poses and a sparse point cloud for you. But then what? The sparse cloud is just a skeleton, not a surface. That's the gap OpenMVS is built to fill.

What It Does

OpenMVS (open Multi-View Stereo reconstruction library) is a computer-vision library aimed at the Multi-View Stereo reconstruction community. It picks up where Structure-from-Motion pipelines leave off. You give it a set of camera poses plus the sparse point cloud, and it gives you back a textured mesh.

The library covers four main stages of the photogrammetry chain:

  • Dense point-cloud reconstruction, to get as complete and accurate a point cloud as possible
  • Mesh reconstruction, to estimate a surface that best explains that point cloud
  • Mesh refinement, to recover fine details
  • Mesh texturing, to compute a sharp, accurate texture for coloring the mesh

It's a C++ library (that's the standard territory for this kind of computer-vision work), and it targets researchers and practitioners rather than casual users. The full documentation lives on the project's wiki, which is worth bookmarking since the README itself is deliberately lean.

Why It's Cool

  • It's honest about being a gap-filler. The README doesn't oversell. It points at mature Structure-from-Motion projects and says plainly that none of them address the last part of the photogrammetry chain. OpenMVS exists to close that loop, and that focus is the whole pitch.

  • The pipeline is complete. Dense reconstruction, meshing, refinement, and texturing are all here. You're not stitching together four different tools with four different file formats—you run one chain from sparse cloud to finished textured mesh.

  • Large scenes are actually handled. This is the detail that stands out. The MvsScalablePipeline.py script runs the full dense-to-texture pipeline on a scene of any size. When the scene won't fit in the memory your machine has available, it splits the scene into sub-scenes, reconstructs each one in turn, and merges the results. A scene that fits gets reconstructed whole. This is the kind of practical engineering that separates a usable tool from a demo—photogrammetry scenes get big fast, and memory is usually where naive pipelines die.

  • Continuous integration across platforms. Windows, Ubuntu, and MacOS x64 builds are all covered by CI, and automatic Windows x64 binaries are posted as artifacts for each commit. If you just want to try it without fighting a build, that's a real convenience.

How to Try It

The README keeps things minimal and points to the wiki for details, so start there. You'll want the building wiki page for platform-specific setup, and the usage example to see how the modules fit together.

A reasonable path:

  1. Clone the repository from github.com/cdcseacave/openMVS.
  2. Follow the building instructions for your platform (Windows, Ubuntu, or MacOS x64). If you're on Windows and don't want to compile, grab the automatic x64 binary from the Artifacts page for any commit.
  3. Work through the usage example on the wiki to understand the input format and module flow.
  4. For a large scene, run the scalable pipeline:
scripts/python/MvsScalablePipeline.py scene.mvs

That single command handles dense reconstruction, mesh generation, refinement, and texturing, splitting the scene automatically when memory runs short. If you're using MvgMvsPipeline.py, note that it does the same for its dense steps by default—pass --sub-scene-memory off if you want to disable the behavior.

Final Thoughts

OpenMVS isn't trying to be everything. It's a focused library for people who've already got camera poses and sparse geometry and need the rest of the pipeline—the dense, messy, memory-hungry part. If you're a computer-vision scientist or a photogrammetry practitioner, this is a sensible piece of infrastructure to know about, especially if your scenes are large enough to strain your RAM. The README is sparse, but the wiki and the scalable pipeline script tell you the project takes real-world usage seriously. Start with the example, then scale up.


Follow @githubprojects for more developer tools and open source projects.

Back to Projects
Project ID: openmvsLast updated: November 18, 2026 at 06:15 AM