Hi Pix4D team,
I would like to request a native COLMAP export in PIX4Dmatic, ideally exposing the existing calibration and sparse point cloud as a standard COLMAP dataset (cameras.txt/images.txt/points3D.txt or the .bin equivalents), together with the option to export the corresponding images (distorted or undistorted).
Background and motivation:
The alignment quality in PIX4Dmatic is excellent, and I would like to reuse it as the basis for 3D Gaussian Splatting. However, the built-in Gaussian splatting output in Matic only exports a finished .ply, with no control over the training itself, as already reported. In practice for example this means I end up with a splat capped at roughly 2 million Gaussians, even though my GPU could handle many times that. There is no way to adjust splat count, iterations, densification thresholds or other training parameters (still).
I have seen that the latest Matic preview version (2.8) added an export for camera positions after alignment/calibration, which is a great step and much appreciated. From my point of view, though, this does not yet help for external 3DGS training in tools like Postshot or LichtFeld Studio, because those pipelines expect a full COLMAP dataset (camera intrinsics with the correct model, per-image extrinsics, and a sparse point cloud for initialization), not just camera positions.
As a result, the workaround today is still to redo the entire alignment and georeferencing in another tool (for example RealityScan), export COLMAP from there, and feed that into the external trainer. This feels unnecessary: the bundle adjustment is already done in Matic, and to a very high standard. It seems redundant to recompute the same alignment in a second piece of software just to obtain a COLMAP export. If Matic already holds the full solved geometry, why not let users export it directly in the common COLMAP format?
If Matic could export a COLMAP dataset directly from an existing calibration, I could hand the Matic alignment straight to any external 3DGS trainer and control the output quality myself. This would:
- let users bypass the limitations of the built-in splatting output
- reuse Matic’s high-quality alignment instead of recomputing it elsewhere
- integrate Matic cleanly into modern radiance-field pipelines (LichtFeld Studio, gsplat, Postshot, Nerfstudio, etc.), which all expect COLMAP as input
Nice-to-haves for such an export:
- choice of camera model (OPENCV / FULL_OPENCV / PINHOLE)
- toggle for distorted vs undistorted image export
- option to include the sparse point cloud as points3D for initialization
- standard COLMAP directory layout (sparse/0 + images)
For reference, RealityScan already offers exactly this kind of COLMAP export and it plugs directly into external trainers. Building on the camera export you already added in 2.8, extending it to a full COLMAP dataset would make Matic a much stronger front end for high-fidelity Gaussian splatting.
Thanks for considering this.
Manuel