You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
I'd like to discuss a feature concept so I can get it out of my head, and get on with the ROS2 port.
I'm working with Fuse and GTsam, and I keep seeing a limitation in the expression of state variables.
Constraints are a function that creates a vector of residuals,
Variables are data structures that hold a parametrisation of state variables.
Ceres is fed cost functions that take Variables' values directly, and an automatic derivative of the measurement characteristics as provided by the Constraint.
This means that if you want to change the nature of a Variable, you also need to rewrite all of your Constraints.
In a hypothetical deep refactor in the distant not-now, Would it be reasonable to modify the Variable so that it exposed measurable properties that are a function of its parametrisation instead of exposing the parameters directly?
This would allow a Constraint to be defined over, the distance between two 2d points at some as normal, but the Variable would be able to hold the control points of an interpolated trajectory and calculate a 2d point representing the location at the time of measurement, and have the relationship with the control points be fed through Ceres auto-differentiation with the rest of the Cost Function
Seeking discussion only, I don't intend to add this feature in a hurry
The text was updated successfully, but these errors were encountered:
I'd like to discuss a feature concept so I can get it out of my head, and get on with the ROS2 port.
I'm working with Fuse and GTsam, and I keep seeing a limitation in the expression of state variables.
Constraints are a function that creates a vector of residuals,
Variables are data structures that hold a parametrisation of state variables.
Ceres is fed cost functions that take Variables' values directly, and an automatic derivative of the measurement characteristics as provided by the Constraint.
This means that if you want to change the nature of a Variable, you also need to rewrite all of your Constraints.
In a hypothetical deep refactor in the distant not-now, Would it be reasonable to modify the Variable so that it exposed measurable properties that are a function of its parametrisation instead of exposing the parameters directly?
This would allow a Constraint to be defined over, the distance between two 2d points at some as normal, but the Variable would be able to hold the control points of an interpolated trajectory and calculate a 2d point representing the location at the time of measurement, and have the relationship with the control points be fed through Ceres auto-differentiation with the rest of the Cost Function
Seeking discussion only, I don't intend to add this feature in a hurry
The text was updated successfully, but these errors were encountered: