WIP constraints with precision matrix #5
Conversation
…se in creating asymetric constraints
…eedless covariance
…e weight on low variance directions, and vice versa
catskul
force-pushed
the
feat-precision-matrix-for-constraints
branch
from
4face2d to
af3711a
Compare
…ass through to use choleskey instead of full precision
|
About to push a bunch of changes so might make sense to wait on review. |
|
I'll have to clean up that history before cartographer people would accept. |
|
I remembered that cartographer used to have a weight matrix but dumped it a long time ago. It appears they were using if or a different purpose though the original paper appears to endorse the use of covariance in the constraint for improved accuracy. https://github.com/googlecartographer/cartographer/pull/309/files |
| auto row_major_jacobian = Eigen::Matrix<double, Eigen::Dynamic, 3, Eigen::RowMajor>(point_cloud.size(), 4); | ||
| auto jacobianPtr = row_major_jacobian.data(); // for use to create fake array of jacobians | ||
|
|
||
| costFunction->Evaluate(parameter_blocks.data(), residuals.data(), &jacobianPtr); |
There was a problem hiding this comment.
Not of any concern here, but there is also problem.Evaluate method that can populate residuals and jacobians using the final solution. FYI
There was a problem hiding this comment.
I didn't use that one because the problem still has the seed in it. As far as I've read removing a cost function can mess up the jacobians.
| ceres::Solver::Summary unused_summary; | ||
| Eigen::Matrix3d seedless_precision; |
There was a problem hiding this comment.
May consider initializing to zeros here. Ceres also uses sparse matrices(sometimes) and depending on the implementation, there may be a condition in which the Evaluate may only assign to the non-zero entries of the matrix leaving garbage value for the zero entries.
No description provided.