search

ethz-asl / aslam_optimizer

Optimizer classes for aslam_cv, kalibr, aslam_incremental_calibration, ..
Other
35 stars 7 forks source link

When using an M-Estimator, the wrong cost is used #11

Open furgalep opened 10 years ago

furgalep commented 10 years ago

We are not using \rho(x) as the cost. We are using w(x) x^T x.

This should be changed.

furgalep commented 10 years ago

See here for a possibility of how we could implement M-Estimators. They have to return p(x), p'(x), p''(x).

http://homes.cs.washington.edu/~sagarwal/ceres-solver/stable/modeling.html#LossFunction

furgalep commented 10 years ago

Here is where the M-Estimator is applied:

https://github.com/ethz-asl/aslam_optimizer/blob/master/aslam_backend/src/ErrorTerm.cpp#L20-L24

furgalep commented 10 years ago

These also need attention:

https://github.com/ethz-asl/aslam_optimizer/blob/master/aslam_backend/include/aslam/backend/implementation/ErrorTerm.hpp#L169-L192

furgalep commented 10 years ago

And

https://github.com/ethz-asl/aslam_optimizer/blob/master/aslam_backend/src/ErrorTermDs.cpp#L144-L167

furgalep commented 10 years ago

Maybe other places as well. Thanks!