Expose a bundle_adjustment function

[blurgyy]

This commit exposes a bundle_adjustment function that has a similar functionality to the command line utility colmap bundle_adjuster. One use case it enables is to allow one use pycolmap to run bundle adjustment after sparse reconstruction to refine the principal point, as jointly refining the principal point¹ during sparse reconstruction would make sparse reconstruction on some scenes to fail.

Note that not all available members of ceres::Solver::Options are exposed in this change, but only the ones modified by BundleAdjustmentOptions in colmap². Besides, the registered CeresSolverOptions uses py::module_local() to avoid conflicts with downstream libraries.

Example usage

from pathlib import Path

import pycolmap

sparse_reconstruction_directory = Path("sparse/0")  # or just `... = "sparse/0"`

ba_opts = {
  "refine_principal_point": True,
  "solver_options": {
    "max_num_iterations": 256,
  },
}

pycolmap.bundle_adjustment(
  input_path=sparse_reconstruction_directory,
  output_path=sparse_reconstruction_directory,
  options=ba_opts,
)

Known problem

Though creating a pycolmap.BundleAdjustmentOptions object with a nested dict works as expected, i.e., the following examples work well:

import pycolmap

ba_opts = pycolmap.BundleAdjustmentOptions({
  "refine_principal_point": True,
  "solver_options": {
    "max_num_iterations": 256,
  },
})
print(ba_opts.todict())  # runs successfully with user-specified option values

ba_opts = pycolmap.BundleAdjustmentOptions({
  refine_principal_point=True,
  solver_options={"max_num_iterations": 256},
})
print(ba_opts.todict())  # runs successfully with user-specified option values

it gives an error while trying to create a
pycolmap.BundleAdjustmentOptions object with its solver_options set to a pycolmap.CeresSolverOptions object, i.e. the following gives an error:

import pycolmap

solver_options = pycolmap.CeresSolverOptions(max_num_iterations=256)
print(solver_options.todict())  # runs successfully with user-specified option values

ba_opts = pycolmap.BundleAdjustmentOptions({
  "refine_principal_point": True,
  "solver_options": solver_options
})  # RuntimeError: instance allocation failed: new instance has no pybind11-registered base types

The "new instance" mentioned in the RuntimeError message probably points to the CeresSolverOptions added in this commit, but I'm not sure about a way to solve this at the moment, plus, as reported above, using dicts to specify the option values works properly and works around this error.

Footnotes

1. https://colmap.github.io/faq.html#frequently-asked-questions ↩

2. https://github.com/colmap/colmap/blob/3.8/src/optim/bundle_adjustment.h#L81-L89 ↩

[sarlinpe]

Thanks! This is most likely a limitation of make_dataclass, we will look into it.