COLMAP 3.12 is here, and it is particularly good news for anyone turning their photos or videos into radiance field reconstructions such as Gaussian Splatting scenes. Released today, the update touches every stage of the structure from motion pipeline, from image matching to final bundle adjustment, and many of the improvements translate directly into shorter training times, cleaner geometry, and more reliable camera poses.
The headline feature is native support for sensor rigs. Instead of shoe horning multi camera arrays or rotating panoramas into single lens assumptions, you can now describe each sensor’s offset inside a shared rig and let COLMAP propagate those relationships during reconstruction. If you capture 360 degree panoramas for Immersive Gaussian Splatting or run a GoPro ring around a product, the new rig should be much more capable. The pose output is now split between rigs and frames files, so downstream code that reads images.txt still works, while new code can dive into rig metadata for finer control.
Geo-referencing is simpler too. Built in conversion between latitude longitude altitude and UTM coordinates, together with a viewer that honors arbitrary origins, lets you inspect drone captures in absolute GPS space before handing them to your neural renderer. Large outdoor splats benefit because scale drift and origin shifts are visible immediately rather than halfway through training.
Under the hood, the solver stack has been tightened. Minimal solvers for affine and generalized pose now run faster and expose Python bindings and absolute pose refinement minimizes pixel space error for better fisheye performance.
There are a few caveats. Rig information lives in new database tables and additional reconstruction files, poses are composed as “sensor-from-rig times rig-from-world” rather than a single matrix, and the dedicated rig bundle adjuster has been folded into the default optimizer.
Ubuntu-20.04, macOS-13, and Visual Studio 2019 drop out of official support, though they may still compile with effort. Make sure that you check if you have any dependencies that might be using these, especially with Visual Studio 2019 and Nerfstudio.
Several bug fixes rounds out the release: two view pose estimation is more stable, covariance propagation is correct for both absolute and relative poses, and a rare deadlock in vocabulary-tree matching is gone. Minor annoyances in the GUI, bitmap loader, and CUDA builds have also been ironed out.
COLMAP comes with a BSD license and the full change log can be accessed here.
New to Gaussian Splatting? Start here
