Comparison to other non linear least square solver

[TheFrenchLeaf]

Hi,

Have you done some short comparative study with the following package ?

sba http://www.ics.forth.gr/~lourakis/sba/

g2o http://openslam.org/g2o.html

MCBA/PBA http://grail.cs.washington.edu/projects/mcba/

I'm really impatient to test it.

Regards,

Pierre

[Sameer Agarwal]

Pierre,

We have not done any comparisons. That said, there is a good chance that for small to medium sized problems these solvers will beat Ceres easily. Two reasons.

1. Ceres does a bunch of stuff to make it a general purpose solver, which has some overhead.

2. Ceres deals with bundle adjustment problems with arbitrary sparsity (any camera can share anything with any other camera), which means its linear solver is much more complicated, so it can't take some of the shortcuts these other solvers take.

That said, it also has a few things that other solvers don't. It has multithreading for jacobian and schur elimination, and it has autodiff. 

We would be very curious to see head to head numbers, but we have not had a chance to do that comparison ourselves.

Sameer

--
----------------------------------------
Ceres Solver Google Group
http://groups.google.com/group/ceres-solver?hl=en?hl=en