We wanted to share our first results using AliceVision, a free and open source photogrammetry framework.
AliceVision is the result of a synergy between multiple Universities & Labs (Czech Technical University, IMAGINE, Institut National Polytechnique de Toulouse (INPT), Simula Research Laboratory, and Quine) and Mikros Image, a French post-production company. The team started a project called OpenMVG a few years ago. They presented the new AliceVision Framework at FMX 2018.
Here is a workflow video presentation, done by the AliceVision team :
First Test: Results are Promising!
We ran a quick test using an old dataset of 35 photos (Facial Capture Rig, using Canon 100D DSLRs). Here is the raw reconstruction result using full-size depth maps generation. We compare it with the result obtained with a great commercial photogrammetry solution: Reality Capture by CapturingReality:
RealityCapture mesh is obviously sharper and less noisy, but the AliceVision result is very promising! There are lots of options to tweak & test for depth maps generation and mesh refinement steps.
What We Liked About AliceVision
Free & Open Source
We love the open source philosophy, and we hope it will help the software to grow faster. If you wan’t to try the software and contribute, you can go to AliceVision Github Profile.
Beautiful Python GUI: Meshroom
Core AliceVision Framework consists of different programs and libraries but does not ship with any Graphical User Interface as it is. The team developed a nice GUI named Meshroom to interface with the library.
Meshroom is written is Python (Compatible with Python 2 & Python 3) using modern QML features of PySide2. It’s cross-platform and really easy to install and tweak up to your needs.
We love that the UI is separated from the standalone core framework, it makes it super-easy to write custom user interfaces, for example, a set of nodes in SideFX Houdini to run the different steps of the reconstruction.
Designed for production & Using Standard file formats
Meshroom / AliceVision have a lot of production-oriented features: nodal Interface, out-of-the-box command line interface, clean logs. It’s really easy to write submitters for computing farms or even for running reconstructions on a cloud-based solution like AWS or Google Cloud.
A really nice point is that Alicevision uses standard open source file formats for processing. For example, structure from motion result is stored as an Alembic scene (with cameras and point cloud), depth maps are stored as OpenEXR files etc… That makes it easy to use the results directly in other 3D softwares without running any “export” or conversion step.
You can for example directly open and use the depthmaps using any software able to import OpenEXR files, or you can directly preview them in Meshroom.
Quick Compare With Commercial Solutions
Here is a comparative test between AliceVision and two other great (Commercial) photogrammetry solutions: Agisoft Photoscan and CapturingReality Reality Capture. We use the Vase MVS Dataset from Stanford for these tests. Detected Feature number per image was about ~3K in each software (pretty low resolution dataset). We ran full-resolution depth map estimation (High Quality in RealityCapture, Ultra-High in Photoscan, and x1 in AliceVision)
The following graph shows the time spent during structure from motion (feature detection + alignment), depth maps computing + meshing, and texturing phases. Please note that the results presented in the graph are not very precise, we tried to use similar settings in each software just to give you a quick estimate, take it for what it’s worth.
And here are the results for each software (mesh + texture) :
No surprise, in terms of speed and mesh quality Reality Capture wins the race followed by Photoscan and then AliceVision. Note that the photos dataset lacked metadata (focal length, camera model …), and these pieces of information seem pretty important for AliceVsion to give an optimal alignment.
The results are still encouraging for Alice vision. It would be more “fair” to compare AliceVision with other open source photogrammetry solutions, keep in mind Reality capture and Photoscan are both (great) commercial softwares, with multiple years of existence and considered as the best solutions by many people in the industry. We hope that people will contribute to the AliceVision project to make it reach Photoscan/Reality Capture performances, it’s actually not that far !
To learn more about AliceVision, you can check the official website: https://alicevision.github.io/
If you want to contribute, here’s the AliceVision Github page: https://github.com/alicevision
We have successfully built the Framework on Windows 10 / VC140. Official binaries should be planned, but if you are interested in early Windows binaries for testing, you can drop a comment here or on the post of our Facebook Page. If many people are interested, we’ll consider releasing early unofficial Windows binaries.
Edit : Official binaries (Win/Linux) are available! https://github.com/alicevision/meshroom/releases/tag/v2018.1.0
Thanks for reading, if you liked this article, let us know 🙂 !