Meshroom

If you want to turn real-world objects into 3D models, you will notice a common problem - first, the 3D scanners are expensive and second, the modelling software requires artistic skills.

Meshroom can help you create 3D models from regular photos. It does this by analysing overlapping features in your images and it calculates the camera positions. You take photos around your object, feed them in the software, and you will obtain a textured 3D mesh out.

You can use Meshroom to digitise museum artefacts, or create game assets from real objects, or preserve architectural details etc. The software was created as a part of the AliceVision framework - which is a photogrammetric computer vision project developed by multiple universities and labs.

Important: Please note that it requires an NVIDIA GPU with CUDA compute capability 3.0+ for full functionality. Without a compatible graphics card, you're limited to Draft Meshing mode (it is slower and has a lower quality).

Meshroom Features

Feature	Description
Node-based pipeline	You can see every processing step as a visual node and you can modify the parameters
Auto process	You can drag photos and press start - the software will handle everything
Add photos mid-project	You can add more photos to existing reconstructions - if you find gaps
Multi export formats	You can output OBJ files with textures for some third-party apps such as Blender, Unity, Unreal
Draft meshing mode	You can process without using a NVIDIA GPU it will use only the CPU
Sketchfab integration	You can upload models directly to Sketchfab
Camera database	Meshroom will recognise most camera models automatically for better calibration
GPU acceleration	If you have CUDA-enabled graphics card you will get faster processing
Project portability	You can move your projects between computers using the .mg file and cache folder
Live preview	You can see the results as each processing stage completes
Multi pipelines	You can run different settings at the same time for comparison
Batch available	You can start multiple projects to run overnight

Meshroom Interface

You will notice that Meshroom uses a split layout (it is similar to other node-based apps)

The left panel will show the imported photos as thumbnails. The central panel will show you the 3D viewer where you can see the reconstruction progress. The bottom will show you the node graph - the entire process. The right panel will show you the image viewers and properties.

The Meshroom interface is quite technical, I would say quite intimidating when you see all the nodes. But the good thing is that you don't need to understand them.

The photo import is also easy and you can drag images into the left panel - there are no complicated import dialogs.

The 3D viewer also works great. You can rotate the model and zoom.

The process operations are quite intensive. If you add a lot of photos (for example 100 or more) it will take you an hour or more. The more you add, the more time it will take (quite obvious). If you need to use the "Draft Meshing" (without NVIDIA GPU), it will work but it will take longer and you will get lower quality results.

You might be disappointed about the low quality results when you use the "Draft Meshing" but if this is important at least you know what to expect before investing more time or money.

Good to know: You can save your project at any time and resume it later. Actually, I would recommend you to save it immediately after importing photos - always. Since processing can take hours, it is a good thing to protect against crashes - always. If you're an old Windows user like me (since Windows 95 etc. you know this is a great idea).

Installation is different - I also made a quick table for each operating system (Windows seems to win here).

OS	Instructions
Windows	You can download the ZIP file, extract it anywhere, and run Meshroom.exe directly. There is no setup or install. I tested the latest version on a Windows 11 machine with an RTX 3060, and it runs great. Please make sure you have the updated NVIDIA drivers.
macOS	Meshroom will require some extra steps if you are using a Mac. You will need to install Meshroom via the command line or you can use a package manager.
Linux	The AppImage package should work on most distributions. Just download the "AppImage", make it executable and run it.

Note for the Linux users: This is quite important if you are using Linux: To make it executable in your terminal, try this command:

chmod +x Meshroom*.AppImage

Note for the macOS users: Also, the GPU acceleration might not work depending on your Mac model.

A quick preview of the Meshroom Process

I tried to document this so you can understand better what's going on. Each step has a quick description.

Step	Name	Description
1	CameraInit	This will read your photo metadata: it will load image metadata, sensor information and generate viewpoints etc.)
2	FeatureExtraction	It will find distinctive points in each image: it will extract features from images, including the descriptors for those features, change the file extension depending on what feature you're extracting.
3	ImageMatching	It will identify which images see the same parts: a pre-processing step with an algo that will attempt to match images
4	FeatureMatching	It will match features between paired images: it uses the feature descriptors to find the correspondence between images
5	StructureFromMotion	It will calculate the camera positions (this is when you first see results): it will reconstruct the 3D points from the input images
6	PrepareDenseScene	It will prepare images for depth calculation: node undistorts the images and it will generate EXR images
7	DepthMap	It will compute depth for each pixel (but this node requires NVIDIA CUDA GPU)
8	DepthMapFilter	It will remove inconsistent depth values: please note that this is not 100% consistent. Some depth maps will claim areas occluded by other depth maps.So this step will isolate these areas and it forces depth consistency.
9	Meshing	It will create the 3D surface: it will generate Mesh from SfM point cloud or from DepthMap
10	MeshFiltering	It will clean and smooth the mesh: it will filter the unwanted elements of from your mesh
11	Texturing	Finally, it projects photos onto the mesh: Texturing will create UVs and projects the textures change quality and size or file type of texture

Keep in mind that each stage will show you a coloured progress bar.

Green means complete, orange means processing, blue means waiting.

If there is a failure, you will see the red colour and you can check the log for more details.

A few Tips for Better Results

I tested a few models with Meshroom, and here's what worked for me:

Take at least 50 photos - 30 is the minimum but the more pics you take - the better the final result. Walk around your object slowly, overlap each shot by 60-70%. Try to avoid any shiny or transparent surfaces (they will confuse the algorithms). The textured surfaces will work much better than plain colours.

If you're using a DSLR, set the aperture to f/5.6 or higher to keep everything in focus. Try to shoot in good, even lighting - harsh shadows will cause you problems. Try not to use the flash unless you can keep it consistent.

For small objects, it is great if you can use a turntable and rotate the object instead of walking around. For buildings, you can take photos at different heights - ground level, eye level, and if possible from above (hard and you need a drone).

Meshroom Alternatives

Based on my research and testing, there are a few Meshroom alternatives:

Name	Free or Paid	The main difference
Agisoft Metashape	Paid - an expensive license)	It is expensive but you get PRO features, masking tools, faster processing, support etc.
RealityCapture	Paid (PPI model)	It is probably the Fastest processing option but it has a pay-per-project license type
COLMAP	Free	It is command-line focused, therefore not for beginners, more for the advanced users but faster than Meshroom
3DF Zephyr	Free/Paid versions	It has a user-friendly interface but also a 50-photo limit on the free version
WebODM	Free/Paid cloud	This one is designed for drone mapping, so great for aerial photogrammetry

Just tossing out a few ideas regarding the Meshroom alternatives

Metashape is probably the closest commercial alternative but it is expensive if you need the Pro version. RealityCapture has the best speed and quality but the pricing model might be expensive. It could work if you need someething once in a while or just a single time.

COLMAP as I said it will also offer you free photogrammetry but it requires technical knowledge and command-line skills. 3DF Zephyr is somewhere in the middle with its free version but there is a limit - up to 50 photos. WebODM is focused on the drone imagery rather than object scanning.

Final Conclusion

You can use Meshroom and obtain PRO 3D reconstructions for free. For example, you can use it to create game assets or preserve historical artefacts, or generate models for 3D printing. It can process the entire photogrammetry pipeline from photos to textured mesh

The best thing is that the AliceVision team worked hard and kept it updated over the years. Yes, the NVIDIA GPU requirement might look like a limitation (it isn't in my opinion and I could make an analogy with many things - two different cars etc.) Yes, the processing time will test your patience - I've had reconstructions run overnight. But the final result will compete with a similar software that will cost you a lot of money.

If you check the alternatives, Meshroom is great, and if you have the required hardware and time, it is actually amazing. If you like Meshroom, please consider a donation to the AliceVision project for this great free software.

Full release notes on GitHub: https://github.com/alicevision/meshroom/releases/tag/v2023.3.0