Re: [eigen] New Levenberg Marquardt stuff

[Christoph Hertzberg <chtz@xxxxxxxxxxxxxxxxxxxxxxxx>]

On 07.12.2012 23:44, Victor Liu wrote:
> [moved your response to the bottom]
> On 12/07/2012 02:36 PM, Thomas Capricelli wrote:
> > Il 07/12/2012 23:18, Sameer Agarwal ha scritto:
> > > Eigen is a linear algebra library (or atleast thats how most people
> > > think about it).
> > > why should a full fledged nonlinear least squares library be included
> > > in it?
> >
> > Sure it would be insane to include the "full-fledged" ceres in eigen.
> > But i disagree that some LM implementation has nothing to do in eigen :
> > as most nonlinear algorithm, it's mostly "do a lot of linear operations
> > in loops". Very much in the topic of eigen.
> I have often thought about using Eigen for my own scientific computing
> projects, but each time I have been deterred by what I consider its
> enormous size (don't even get me started about Boost). To include a
> nonlinear optimizer in it would just make things that much worse. I know
> there are a lot of projects out there that use Eigen only for fixed size
> 3x3 or 4x4 matrices for use in graphics, or some that use it for basic
> matrix multiplication/inversion and such. Including LM would be far
> outside the realm of relevance for them.

I agree that putting a full-fledged non-linear LS solver into Eigen is 
way out-of-scope.
And implementing "just Levenberg-Marquardt" within Eigen has a quite 
limited use IMHO, as I think that for all non-trivial optimization 
problems there are still the tasks of variable management and 
calculating the Jacobian.

Actually, I would suggest starting a side-project for non-linear least 
squares problems, joining the efforts of all existing algorithms -- 
having a nice Eigen-like API using the most efficient solvers under the 
hood. I might write a separate mail regarding this to the authors of 
g2o, iSAM and ceres the next days (if someone else wants to join, just 
tell so)


Christoph


--
----------------------------------------------
Dipl.-Inf. Christoph Hertzberg
Cartesium 0.049
Universität Bremen
Enrique-Schmidt-Straße 5
28359 Bremen

Tel: +49 (421) 218-64252
----------------------------------------------