Jianyuan Wang

true I am "encouraging" some people to build something like stereoanything, hope they make it fast enough lol

It seems there is no need to run whole sfm with known camera intrinsic and extrinsic. You just need to run BA on it. Check https://github.com/colmap/pycolmap/blob/b6627db2266f098c21c3d7e4b7844b4b90d8e02d/pycolmap/pipeline/sfm.h#L95

you can check https://github.com/colmap/colmap/blob/393fbc3fb9e55fc8bc0027ac237a0cf62b461933/src/colmap/controllers/incremental_mapper.cc#L38 or https://github.com/cvg/pixel-perfect-sfm You can also try https://vggsfm.github.io/ but I have not released the code for it.

Hey @1ssb, we have used COLMAP on Re10k and the result is relatively sparse. You can check Figure 6 in https://openaccess.thecvf.com/content/ICCV2023/papers/Wang_PoseDiffusion_Solving_Pose_Estimation_via_Diffusion-aided_Bundle_Adjustment_ICCV_2023_paper.pdf for how sparse the point cloud it is. We...

well I did not try many solutions because it was not an important part of our paper (we just did it to visualise the point clouds in re10k), but yes...

The simplest solution may be: Run on all images -> Record wrong registrated images -> Filter them out-> Run COLMAP again on filtered images. If you don't want to run...

I think this is what you want https://github.com/PoseLib/PoseLib if else, you can try to bind essential_matrix_estimation by yourself, e.g., like what I have done in: https://github.com/colmap/pycolmap/pull/252

The short answer is using conda. Most libraries required by colmap (e.g., in https://colmap.github.io/install.html) is provided in conda especially conda-forge. I successfully compiled it on different linux clusters without root.

If you are happy with python, I would suggest to use SP+SG as in https://github.com/cvg/Hierarchical-Localization/.

Have you installed python3-dev package? It seems python Interpreter cannot be found correctly. If it still does not work after installing python3-dev, you can also consider downgrade to python 3.9.