[Photogrammetry] Meshroom 2023.1 released

I’ve had a keen eye on Alicevision Meshroom since it’s first major release back in… 2018?

Development of the open source photogrammetry software has felt slow for the past few years, but just this week a new release was pushed out, 2023.1.0.

Here’s the official new feature list:

Release Notes Summary

• Major improvements of the depth map quality, performances and scalability. The full resolution can now be computed on most of the standard GPUs.
• FeatureExtraction is now using DSP-SIFT by default for the 3D Reconstruction pipeline.
• Capacity to create panoramas with very high resolutions using a limited amount of memory.
• Enhanced interpretation of RAW images, including new support for Adobe Digital Camera Profile and Lens Camera Profiles databases (if installed on your workstation).
• Improved color management with OCIO support and more options to export in various colorspaces including ACEScg.
• New graph templates enabling users to create custom pipelines.
• Expose a new experimental pipeline for Camera Tracking.
• Improved GraphEditor with copy-paste and multi-selection.
• Improved ImageGallery with thumbnails cache and search options.
• 2D Viewer is now using floating-point images by default.
• And a very large amount of UI improvements and bug fixes.

The full list is available here: https://github.com/alicevision/meshroom/blob/v2023.1.0/CHANGES.md

For my workflow, the changes are pretty minimal. The core software still requires and Nvidia card and uses CUDA (but see Meshroom-CL). While I’ve changed computers since I last tested it, it’s still on the slower side – my Surface Laptop Studio with Nvidia A2000 plus an eGPU with an Nvidia 3070 took over 25 minutes on my 53 image dataset, with optimized settings which is… well, it’s not great, given Reality Capture takes just 2 and a half minutes.

Dissapointingly, while I made sure DepthMap was using all available nvidia GPUs, and it correctly identified them in the log, task manager showed almost zero usage:

The final model was pretty nice though:

It’s worth noting, that only a couple of modules actually use the Nvidia GPUs, according to task manager, which is a real shame.

The interface feels good, though I notice that if you’re not using 100% image scaling in windows (I’m using 125%) then you can’t actually click on a lot of icons because the window size adjustment bars overlap, which is frustrating.

Texturing seems to be fixed, in that ABF and LSCM don’t crash the software anymore.

Hovering your mouse over an image tells you RGB and Alpha values:

But there doesn’t seem to be anything in particular to do with this info, say setting a mask.

A big feature I was waiting for, aside from just overall better performance, was being able to easily scale and transform models, ideally by clicking points in images, but as far as I can tell, this is just not available.

Conclusion

So overall, it’s not a game-changing release, just a few nice small updates. But, as far as I’m concerned it’s still the best and most powerful free and opensource photogrammetry software available. It’s just that given Reality Capture has free educational licences, there’s little reason for me to be using Meshroom these days.