This project contains the implementation of a serial multigrid solver for bundle adjustment in addition to a distributed memory Levenberg-Marquardt implementation in PETSc.
Bundle adjustment problems are stored in either the Bundle Adjustment in the Large format (http://grail.cs.washington.edu/projects/bal/), or in a binary version of it.
- BALUtils: julia code for interacting with bundle adjustment problems.
- ba-problems: test data and results.
- ba-tao: Distributed memory bundle adjuster. Also contains the test harness for the serial solver.
- ba_eig: tools for computing the eigenvalues/eigenvectors of bundle adjustment problems.
- bamg: the multigrid solver (written in Julia).
- ceres-solver: a fork of Ceres Solver with hooks to use the serial multigrid solver.
- city2ba: tools for generating synthetic problems.
- models: 3D city models used to generate synthetic problems.
- multigrid_analysis: tools to apply compatible relaxation analysis to the multigrid solver.
- notebooks: jupyter notebooks for result analysis and plotting.
- options: various configuration used for testing bundle adjustment solvers.
- petsc: fork of PETSc with a parallel Levenberg-Marquardt implementation.
This project requires:
With these installed, run setup.bash to build everything.
Build.hs is a shake build file that can build everything and run tests. It requires ghc (the Haskell compiler) and the shake package to be installed (you can install it with cabal install shake). There are a variety of targets that it can build:
ba-problems/*/ceres_benchmark_pbjacobi.csv: collect results for the point block Jacobi preconditioner.ba-problems/*/ceres_benchmark_visibility.csv: collect results for the visibility preconditioner.ba-problems/*/ceres_benchmark_multigrid.csv: collect results for the multigrid preconditioner.
You can list all targets by running ./build.sh -h.
The ba-problems directory contains test datasets. Unfortunately they are too large to store on github. Contact me if you want access to them.