Function optimizer.computeMarginals()

[andelby]

hi, the result of the optimizer.computeMarginals(), is the covariance matrix about the estimated pose? https://github.com/jingpang/LearnVIORB/blob/b9e73cefee88d7ab352508e1e11aab0990195e91/src/Optimizer.cc#L3745

[ferreram]

Hi, the function computeMarginals() returns the Hessian matrix associated to the state vertex(0), you have to invert it to get the covariance.

[andelby]

1)thanks @ferreram for the reply, I tried writing this code: g2o::SparseBlockMatrixXd spinv; optimizer.computeMarginals(spinv, optimizer.vertex(0)); Eigen::Matrix<double, 6, 6> margCovInv = Eigen::Matrix<double, 6, 6>::Zero(); margCovInv.topLeftCorner(6,6) = spinv.block(0, 0)->inverse(); cout<<"Matrix covariance:\n"<<margCovInv<<endl;

But the matrix seems to be incorrect (values too high?)

Did I miss something in the code?

2)This covariance matrix refer only to the last pose, or to the full trajectory?

[ferreram]

1) Well, the values of the Hessian matrix come from this expression: J' W J, where J is the Jacobian of your optimization problem and W the information matrix (the inverse of the covariances given to each measurements).

The resulting Hessian matrix hence depends on the optimization problem you have defined. The fact that you get high values for the state vertex(0) means that you have a high confidence in the estimated parameters (if you invert this Hessian matrix you will get the covariance matrix associated to the estimated state).

2) I think that in g2o, the computeMarginals() function simply provides the block in the Hessian matrix associated to the state you require (but I don't remember well, you should check that). If it is indeed done like this, then the matrix that you get correspond to the confidence you have in vertex(0) according to the optimization problem you have defined. Hence, if you have performed a Global Bundle Adjustment like optimization, the matrix you have obtained correspond to the information matrix related to the states vertex(0) over the full trajectory.

On Fri, Oct 4, 2019 at 1:09 PM Andrea485 notifications@github.com wrote:

1)thanks for the reply, I tried writing this code: g2o::SparseBlockMatrixXd spinv; optimizer.computeMarginals(spinv, optimizer.vertex(0)); Eigen::Matrix<double, 6, 6> margCovInv = Eigen::Matrix<double, 6, 6>::Zero(); margCovInv.topLeftCorner(6,6) = spinv.block(0, 0)->inverse(); cout<<"Matrix covariance:\n"<<margCovInv<<endl;

But the matrix seems to be incorrect (values too high?)

[image: Schermata del 2019-10-04 13-04-30] https://user-images.githubusercontent.com/7497813/66203033-8cf05b80-e6a7-11e9-87b0-2cb1d64bc863.png

Did I miss something in the code?

2)This covariance matrix refer only to the last post, or to the full trajectory?

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/jingpang/LearnVIORB/issues/70?email_source=notifications&email_token=AEPCATTTSXXLCYP3QT6GOJDQM4P5TA5CNFSM4I5MM4W2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEALKDTA#issuecomment-538354124, or mute the thread https://github.com/notifications/unsubscribe-auth/AEPCATQGVWUGA43RXUZTKBTQM4P5TANCNFSM4I5MM4WQ .

[arianna003]

hi @ferreram , the matrix of @Andrea485's figure represent covariance not hessian, but is the pose coordinate uncertainty in meters?

[ferreram]

Hi @arianna003 , yes you're right, my bad I read it too fast. In this case @Andrea485 , the uncertainty values of your covariance matrix indeed seems very high. The uncertainty around the translation parts of the matrix are in squared meters if the transnational parts of the states are expressed in meters (it should be the case with a Visual-Inertial method but not with a monocular SLAM)

[andelby]

@ferreram what's wrong in my code, what can I do?

[ferreram]

@Andrea485 , there is nothing wrong in the code you have used to compute the covariance matrix. But what are you doing? VI-Odometry? If so, is the Bundle Adjustment like optimization is giving you good results? You can also have a look at the information you are setting for your edges.

As the Hessian matrix is derived from the Jacobian and the information matrix, high covariances will come from high Jacobians value. Maybe you should verify your Jacobian as well (you can test doing automatic differentiation).

[leslieburke]

@Andrea485 I got the same problem, the value of covariance matrix is too high?what wrong with it?