COLMAP Export (new in v1.3)

The COLMAP export packages your reconstruction in the format used across the modern 3D and neural-rendering ecosystem. Replica writes calibrated cameras, poses, a sparse point cloud and lens-undistorted images — everything you need to feed Gaussian Splatting, NeRF and Structure-from-Motion tools without any manual preparation.

Available in all editions — COLMAP export works in Free, Lite and Pro.

Why undistorted images matter

Phone and camera lenses bend straight lines (barrel/pincushion distortion). Most neural-rendering pipelines — and 3D Gaussian Splatting in particular — assume an ideal pinhole camera and cannot work with distorted input: they require undistorted images and the simple PINHOLE camera model.

Replica solves this before export. Each photo is rectified using Apple's per-image lens-distortion data, producing a pure pinhole image where straight lines stay straight. The exported cameras therefore use the PINHOLE model with no k1/k2 distortion coefficients left to estimate — the data is ready to reconstruct as-is.

How to export

1. Open a project with a completed reconstruction.
2. Go to the Export panel.
3. In the Format menu choose Colmap (photos + poses).
4. Click Export and pick a destination folder.

Replica undistorts each photo, converts the poses to the COLMAP convention and writes the package. A diagnostic README is included so you can inspect exactly what was exported.

Optional mesh: if you have Blender installed and its path set in Settings, Replica also converts the model to model.glb and includes it in the package. The COLMAP export itself does not require Blender — only the optional GLB does.

What's in the package

<project>_colmap/
├── sparse/0/
│   ├── cameras.txt      PINHOLE intrinsics: id, width, height, fx, fy, cx, cy
│   ├── images.txt       world-to-camera poses: qw qx qy qz tx ty tz, camera_id, name
│   └── points3D.txt     sparse colored point cloud: X Y Z R G B
├── images/              rectified (undistorted) photos, matching the PINHOLE intrinsics
├── model.glb            mesh (optional, only if Blender conversion succeeds)
└── README.txt           diagnostics: poses, intrinsics, EXIF orientation, warnings

Conventions

• Camera model: PINHOLE (fx fy cx cy, in pixels). No distortion parameters — the photos are already rectified.
• Pose convention: COLMAP / OpenCV — world-to-camera, +X right, +Y down, +Z forward. Replica converts these from Apple's camera-to-world (ARKit/OpenGL) frame for you.
• Point cloud: the sparse points from the reconstruction, with per-point RGB colour, when available in the project.

What you can do with it

The COLMAP format is the common entry point for the whole reconstruction and neural-rendering ecosystem:

Target	How
3D Gaussian Splatting (3DGS)	Point the trainer at the COLMAP folder. Because images are already undistorted and PINHOLE, you can skip the usual colmap image_undistorter step entirely.
NeRF / Nerfstudio	ns-process-data colmap --data <project>_colmap to build a Nerfstudio dataset, then train (nerfacto, splatfacto, …).
COLMAP	Open the sparse/0 model directly to inspect cameras and points, or continue dense reconstruction.
Blender	Import poses and points with a COLMAP importer add-on for layout, look-dev or reference.
Unreal Engine	Bring the cameras and point cloud into virtual-production and reconstruction pipelines.

Tip — straight to splatting: most Gaussian Splatting repos require undistorted images and a PINHOLE model and otherwise force you to run COLMAP's undistorter first. Replica's package already satisfies both, so the data trains as-is.

Troubleshooting

The README warns "No lens distortion data" for some photos. Replica had no per-image distortion data for those shots and copied the source pixels unchanged. They may not align perfectly if the lens distorts noticeably. This usually happens with images imported from other sources; photos captured directly with Replica carry the distortion data.

The README warns "No intrinsics — wrote a 60° hFoV default". No calibrated intrinsics were available for that image, so Replica assumed a 60° horizontal field of view. Reconstruction tools that refine intrinsics can recover from this, but results are best when real intrinsics are present.

model.glb is missing. The GLB is optional and only produced when Blender is installed and its path is set in Settings. The COLMAP data (cameras, poses, images, points) is exported regardless.

Next Steps

• USD for VFX — the same undistorted photos packaged for camera projection and matchmove
• Workflows (Pro) — automate export as part of a pipeline
• Getting Started — the basics of capturing and reconstructing

Happy reconstructing!