Ceres for curve optimization

[AutoCar]

Hi, 

I am using Ceres to modify a curve (x, y way points) to make it smoother and shorter. So my reside is sum of absolute curvature along the curve and others. My code is as below. I found the optimization result did not move the curve much from the initial value. What option changes can I use to make the optimization more aggressive?

Problem problem;
 DynamicNumericDiffCostFunction<CostFunctor>* cost_function =
        new DynamicNumericDiffCostFunction<CostFunctor>(
            new CostFunctor(num));

    cost_function->AddParameterBlock((int)num);
    cost_function->AddParameterBlock((int)num);
    cost_function->SetNumResiduals(1);

    std::vector<double*> parameter_block = ...
    problem.AddResidualBlock( cost_function, nullptr,  parameter_block);

    Solver::Options options;
    options.linear_solver_type = ceres::DENSE_QR;
   
    //options.minimizer_progress_to_stdout = true;
    Solver::Summary summary;
    Solve(options, &problem, &summary);

[Sameer Agarwal]

two things.

one, why are you solving this as a sum of squares problem if your objective is already the sum of absolute curvature? by squaring the sum you are only making the objective even more non-linear.

two, is the sum of absolute curvature differentiable?

finally you maybe better off using something like the GradientSolver.

also why you using DynamicNumericDiffCostFunction? 

Sameer

--
You received this message because you are subscribed to the Google Groups "Ceres Solver" group.
To unsubscribe from this group and stop receiving emails from it, send an email to ceres-solver...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/ceres-solver/280f63cc-45b8-49e1-8927-9d5b482a46dfn%40googlegroups.com.

[Ge Cong]

User DynamicNumbercDiffCostFunction because curve point number are different from run to run...

To view this discussion on the web visit https://groups.google.com/d/msgid/ceres-solver/CABqdRUD-Fr--rTXR%3DTCkO-Fdy%2BHvd1FGGpvqk06%3DQAbVBUVunw%40mail.gmail.com.

[Sameer Agarwal]

if you are going to use nonlinear least squares maybe try to reduce the sum of squared curvature ? and create one residual per point.

To view this discussion on the web visit https://groups.google.com/d/msgid/ceres-solver/CAES9TMHvR%2BVzYp12T8Mb2ng2MP_-Vq79Hr_mF83Av-5dfGGWKg%40mail.gmail.com.

[Ge Cong]

Hi Sameer:

Could you provide some sample code about how to use Gradient Solver?

THanks,

On Wed, Sep 14, 2022 at 4:27 PM 'Sameer Agarwal' via Ceres Solver <ceres-...@googlegroups.com> wrote:

To view this discussion on the web visit https://groups.google.com/d/msgid/ceres-solver/CABqdRUD-Fr--rTXR%3DTCkO-Fdy%2BHvd1FGGpvqk06%3DQAbVBUVunw%40mail.gmail.com.

[Sameer Agarwal]

https://github.com/ceres-solver/ceres-solver/blob/master/examples/rosenbrock_numeric_diff.cc

To view this discussion on the web visit https://groups.google.com/d/msgid/ceres-solver/CAES9TMH6Z3RDde9iLZ0GG1M0qiVXDX_Cgz%3DLEEu5Op83DnPt%2Bg%40mail.gmail.com.

[Ge Cong]

Thanks, my number of parameters can not be determined at run time...

To view this discussion on the web visit https://groups.google.com/d/msgid/ceres-solver/CABqdRUAA%3DjLxxF6TqtKXBcdM3twuMdzzRC8E135_A%2BsowbHW_g%40mail.gmail.com.

[Sameer Agarwal]

it should be relatively straightforward to hack the numeric differentiation routine to support dynamic number of parameters.

Sameer

To view this discussion on the web visit https://groups.google.com/d/msgid/ceres-solver/CAES9TMFUEvkomhFTFSODgq_XefcH183GbsNO%3Dps-LuC8TRk3_w%40mail.gmail.com.

[Ge Cong]

Seems the version of ceres I have does not have  NumericDiffFirstOrderFunction...

To view this discussion on the web visit https://groups.google.com/d/msgid/ceres-solver/CABqdRUC0pUiK77Ej%2BjGjmkRAz8_zithyPQMgZSBOUgMvio%2BEQA%40mail.gmail.com.

[Ge Cong]

Could you give me some hint how to hack the  numeric differentiation routine to support dynamic number of parameters?

Thank you very much. 

To view this discussion on the web visit https://groups.google.com/d/msgid/ceres-solver/CABqdRUC0pUiK77Ej%2BjGjmkRAz8_zithyPQMgZSBOUgMvio%2BEQA%40mail.gmail.com.

[Sameer Agarwal]

have a look at NumericDiffCostFunction and DynamicNumericDiffCostFunction.

You basically want the latter with just one parameter block.

Sameer

To view this discussion on the web visit https://groups.google.com/d/msgid/ceres-solver/CAES9TMGWBQm%2BVOaOZM%2BvCoVg16PUHM%3DERVV0XGNyy-tCJYkHgA%40mail.gmail.com.

[Sameer Agarwal]

let me see what I can do.
Sameer

To view this discussion on the web visit https://groups.google.com/d/msgid/ceres-solver/CABqdRUCNQ2P_eqUohOKskyVwuNLfjkDS9dObZETzqYymtmtV1Q%40mail.gmail.com.

[Ge Cong]

Could you let me know which version of ceres started to have NumericDiffFirstOrderFunction?

To view this discussion on the web visit https://groups.google.com/d/msgid/ceres-solver/CAK0oyErJ04yqpianfeRM-JL2TvL7OxN8poyxqi-SGXzvRaOugw%40mail.gmail.com.

[Sameer Agarwal]

Try the version of ceres solver at head

with this patch

https://ceres-solver-review.googlesource.com/c/ceres-solver/+/22900

this allows NumericDiffFirstOrderFunction to take runtime number of arguments.

Sameer

To view this discussion on the web visit https://groups.google.com/d/msgid/ceres-solver/CAES9TMHWe58BpXLoD9reco%2B7xVw9ns76vFVtr0yN5jsKz3wf6g%40mail.gmail.com.

[Ge Cong]

Thanks a lot. It worked.

To view this discussion on the web visit https://groups.google.com/d/msgid/ceres-solver/CABqdRUDrvsme9etuXy9b5f5FjHzRuwCXdyjv8L_WKHmcBRrg_Q%40mail.gmail.com.

[Sameer Agarwal]

yw, I will clean the patch up and send it out for review shortly. It should be merged in the next day or two.

Sameer

To view this discussion on the web visit https://groups.google.com/d/msgid/ceres-solver/CAES9TMHr%3DhK7bn8r7u9ZWEd0KM9mw39DUvLvDmXisyUxC9zLpw%40mail.gmail.com.

[Ge Cong]

Thank you for your work. This is really helpful. 

I am wondering if you can also add SetParameterLowerBound to GradientProblem?

Thanks,

To view this discussion on the web visit https://groups.google.com/d/msgid/ceres-solver/CABqdRUA0gdnkgktweNSS9jqZTufY7t9%2B%3D44NZqOp%3DFP9cHMe9w%40mail.gmail.com.

[Sameer Agarwal]

Sorry the first order optimizer does not support bounds constraints yet.

To view this discussion on the web visit https://groups.google.com/d/msgid/ceres-solver/CAES9TMG6%3DB%3Dd%2BB7v_mk4%2B%2BsZSB2Y-77_PV%3Dt7MrRZJMbppX0tg%40mail.gmail.com.