pose_prior_mapper seems like it may be somewhat numerically unstable, or sensitive to initialisation. I have a scene currently which optimises fine saying 10-20% of the time, but otherwise complains about NaNs in the cost function's jacobian every time it runs a global BA step.
Despite this, optimisation seems to progress, but the resulting reconstruction does not look reasonable & experiences scale drift (despite the pose priors).
I'm wondering if this is natural & I should just obtain more priors? Or if perhaps the initial alignment with pose priors should be delayed until more are available or a collinearity check has passed?
Environment:
- OS: Ubuntu 22.04.5
- COLMAP Version 3.11.1, Commit 682ea9a on 2024-12-06 with CUDA)
pose_prior_mapperseems like it may be somewhat numerically unstable, or sensitive to initialisation. I have a scene currently which optimises fine saying 10-20% of the time, but otherwise complains about NaNs in the cost function's jacobian every time it runs a global BA step.Despite this, optimisation seems to progress, but the resulting reconstruction does not look reasonable & experiences scale drift (despite the pose priors).
I'm wondering if this is natural & I should just obtain more priors? Or if perhaps the initial alignment with pose priors should be delayed until more are available or a collinearity check has passed?
Environment: