Ceres Solver error

[swtyree]

Hi all,

I'm getting this error in bundle adjustment:
F20251205 14:33:36.889147 2400324 bundle_adjustment_solver.cc:69] Check failed: solver_options.IsValid(&solver_error) Can't use SPARSE_NORMAL_CHOLESKY sparse_linear_algebra_library_type = CUDA_SPARSE, because support was not enabled when Ceres Solver was built.

Here is the full output (cusfm_error.txt) from running cusfm_cli in case it's helpful. I'm running a clean build of CuSFM on Ubuntu 24 with CUDA 13 and a fresh Conda environment (direct install). Let me know if you need any other info. Thanks!

--Stephen

[JingruiYu]

Hi Stephen,

Thank you for reporting this issue and providing the detailed error log!

From the full output, it looks like the actual root cause may not be directly related to the Ceres Solver configuration itself, but rather stems from earlier stages—most likely poor feature matching or data association. When very few (or no) reliable matches are established between images, the bundle adjustment step ends up with an ill-conditioned or empty problem, which can trigger downstream failures that manifest as solver errors like the one you're seeing.

To help diagnose this, running cusfm_cli with the --enable_debug flag should help. This will save intermediate debugging data (e.g., detected features, matched pairs, etc.), allowing you to inspect whether the feature matching stage is producing sufficient and reliable correspondences.

[brian-xu]

Running into the same error. Running with enable_debug reveals the following issue:

E20260121 00:37:32.967446 32422 visual_vocabulary.cc:730] The built vocabulary contains too few visual words, maybe not mature, occupied ratio:0.0547405,less than 0.7

Any ideas why the vocabulary solver would fail?