I found an issue in COLMAP text export after re-aligning a large project with added control points.
I aligned over 10 000 images from drone + DSLR using control points. Everything looks fine visually, got 1 component with like 98% of images aligned. Then, as usual, tried to export the data to colmap format so I can train a splat. No errors on export, everything saved fine (exported with image undistortion so I have pinhole cameras).
But I cant import this data anywhere, for example to Postshot. It throws an error about wrong database format. So I asked AI to validate the data and it confirmed that:
- Many image records in images.txt reference camera IDs not present in cameras.txt.
- IDs can be very large/unexpected (examples seen in validation: 37566, 79615, 44075, 939468).
- cameras.txt reports around 10,334 cameras, ID range 1..10,334.
- Exported images.txt appears structurally inconsistent in many places (missing keypoint second lines), causing parser desync in downstream tools.
Observed counts from exported files:
- cameras.txt number of cameras: 10,334
- images headers parsed: 10,334
- image->camera missing references: 6,314
- distinct missing camera IDs: 5,153
- image keypoint lines with invalid token grouping (not x y point3D_ID triples): 7,014
Hello @ReverseGravity, can you provide us with detailed repro steps to replicate the problem, which also include the dataset? We will take a look at it.
I’m seeing the exact same issue with a dataset of 4,400 images of a cathedral interior. I used components for the reconstruction.
In RealityScan 2.1, the COLMAP export works perfectly. Using the exact same dataset in 2.2, as well as doing a completely fresh alignment in 2.2, produces corrupted COLMAP TXT output. The sparse points3D.txt is extremely noisy, the geometry is reconstructed incorrectly, and the overall result is unusable.
It really looks like something is broken in the 2.2 COLMAP export.
I could try to upload it somehwere but its like a few gigabytes. Reality Scan creates a separate camera from every image, even if it’s shot with the same lens/settings.. original colmap creates only 1 camera when 1 camera was used.
Anyways, heres what I done:
- imported the drone and DSLR photos, aligned them separately
- created a lot of manual control points to “stitch” them together
- aligned both image sets
- got 1 nice component with like 98% pictures aligned. Everything looked when examined in RS
- selected all cameras in this component and used export → colmap (with image undistortion)
The result was that the .txt files were corrupted. I had thousands more cameras than pictures