Appearance
Meshroom is a free, open-source photogrammetry tool that turns a folder of overlapping photos into a textured 3D model, usually with nothing more than a drag-and-drop and a single click of Start. It is built on the AliceVision research framework, costs nothing, and imposes no export fees — making it the natural first tool for heritage practitioners. The main requirement is an NVIDIA GPU for the dense reconstruction step. This guide walks a complete beginner through one small worked example.
What is Meshroom and what do I need?
Meshroom packages a full photogrammetry pipeline behind a friendly interface. To follow along you need:
- Meshroom (download the latest release for your OS).
- An NVIDIA CUDA GPU for the depth-map stage (the one hard requirement).
- A folder of 40–100 overlapping photos of a single object.
- A few gigabytes of free disk for the cache.
If you have no CUDA GPU, you can run a reduced draft or use a cloud GPU instance; the rest of this guide assumes you have one.
How does the node graph work?
When you open Meshroom you see a node graph along the bottom — a chain of boxes, each doing one job, with data flowing left to right:
text
CameraInit -> FeatureExtraction -> ImageMatching -> FeatureMatching
-> StructureFromMotion -> PrepareDenseScene -> DepthMap
-> DepthMapFilter -> Meshing -> MeshFiltering -> TexturingAs a beginner you do not need to touch any of it. The graph is there so that, later, you can tune or rewire a single stage without rebuilding the whole pipeline.
A small worked example, start to finish
Let us reconstruct a single artefact:
- Open Meshroom.
- Drag your photo folder onto the Images panel on the left. The thumbnails populate
CameraInit. - Press Start. Watch each node turn green as it completes.
- When
StructureFromMotionfinishes, the 3D viewer shows the sparse cloud and camera positions — a quick sanity check. - When
Texturingfinishes, double-click it; the textured mesh appears in the viewer.
On a mid-range GPU, an 80-image set runs in roughly half an hour.
How do I export and open the model?
Meshroom writes results into a cache folder. To get a usable file:
- Right-click the Texturing node and open its output folder.
- You will find an
.objmesh with a companion.mtland texture.png. - Open that OBJ in MeshLab or Blender to inspect, clean, or convert it.
bash
# Quick headless check of the exported mesh in MeshLab
meshlabserver -i texturedMesh.obj -o check.plyKeep the OBJ, MTL, and texture together — moving the OBJ alone loses the texture.
Why did some cameras fail to reconstruct?
Open the StructureFromMotion node and you can see which images aligned and which were dropped. Cameras fail for the familiar reasons: blur, insufficient overlap, or too few shared features. The fix is almost always in the photos, not the software:
| Symptom in SfM | Cause | Action |
|---|---|---|
| Cluster of dropped frames | Coverage gap | Add overlapping shots there |
| Scattered single dropouts | Individual blurry frames | Remove or replace them |
| Whole set won't align | Featureless surface / changing focus | Add texture; lock focus; re-shoot |
Can I improve the result without re-shooting?
Sometimes. If the cameras aligned well, you can push quality by:
- Raising Texturing
textureSideto 8192 for sharper texture. - Increasing Meshing detail for more triangles.
- Recomputing from a higher-quality DepthMap downscale setting.
But if cameras failed to align, no slider rescues them — alignment depends on the photos. Add coverage and reprocess only the affected branch by right-clicking the changed node and choosing to recompute downstream.
How does Meshroom compare to paid tools?
Honestly: for learning and for many heritage objects, Meshroom is genuinely capable. Paid tools like Metashape or RealityCapture earn their cost on speed at scale, scripted reproducibility, and convenience features like easy scale bars. If you are starting out, prove the workflow in Meshroom first — you will understand every stage before paying for anything.
Key Takeaways
- Meshroom is free, open source, and built on AliceVision — no licence or export fees.
- The one real requirement is an NVIDIA CUDA GPU for the depth-map step.
- Beginners can ignore the node graph and simply drag photos in and press Start.
- Export from the Texturing node's output: an OBJ plus its MTL and texture, kept together.
- Failed cameras almost always mean photo problems — gaps, blur, or no features.
- Tune Texturing/Meshing/DepthMap for quality, but only re-shoot fixes alignment failures.
- Learn the whole pipeline in Meshroom before deciding whether paid tools are worth it.
Frequently Asked Questions
Is Meshroom really free?
Yes. Meshroom is free and open source, built on the AliceVision framework, with no licence cost and no export fees. You can use it for personal, research, and most heritage work without paying anything.
Do I need a powerful computer to run Meshroom?
Meshroom's dense reconstruction needs an NVIDIA CUDA GPU for the depth-map step, so a machine without one will be slow or limited. You can still run a draft pipeline that skips the heaviest steps, or use a CUDA-equipped cloud instance.
What is the node graph in Meshroom?
The node graph is a visual pipeline where each box performs one stage, such as feature extraction, matching, or meshing, and arrows pass data between them. Beginners can ignore it and press Start, while advanced users can rewire or tune individual nodes.
How do I export a model from Meshroom?
After processing, right-click the Texturing node and choose to open its output folder, which contains an OBJ mesh with its MTL and texture files. You can load that OBJ directly into MeshLab, Blender, or a viewer.
Why did some of my cameras fail to reconstruct?
Cameras drop out when frames are blurry, lack overlap, or share too few features with their neighbours. Reviewing the StructureFromMotion node shows which images aligned, so you can add coverage where the gaps are.
Can I improve a Meshroom result without re-shooting?
Sometimes: you can raise the Meshing detail, increase texture resolution, or recompute from a better Depth Map setting. But if cameras failed to align, no parameter change will recover them and you must add or replace photos.