This is a big edit but with no templates involed, so it should not be a big deal.
New namespace gtsam::noiseModel collects all noise models, which provide efficient whitening and chain-rule implementation needed for optimization. The class hierarchy gives us the ability to use models from full covariances to i.i.d. unit variance noise with a single interface, where the latter will be much cheaper.
From now on, all non-linear factors take a shared_ptr to a Gaussian noise model. This is done through the parameter (const sharedGaussian& model). The use of a shared pointer allows us to share one noise models for thousands of factors, if applicable.
Just like Richard's Symbol change, there is a compile flag GTSAM_MAGIC_GAUSSIAN which allows you to use doubles, vectors, or matrices to created noise models on the fly. You have to set it to the correct dimension. Use of this is *not* encouraged and the flag will disappear after some good soul fixed all unit tests.
To support faster development *and* better performance Richard and I pushed through a large refactoring of NonlinearFactors.
The following are the biggest changes:
1) NonLinearFactor1 and NonLinearFactor2 are now templated on Config, Key type, and X type, where X is the argument to the measurement function.
2) The measurement itself is no longer kept in the nonlinear factor. Instead, a derived class (see testVSLAMFactor, testNonlinearEquality, testPose3Factor etc...) has to implement a function to compute the errors, "evaluateErrors". Instead of (h(x)-z), it needs to return (z-h(x)), so Ax-b is an approximation of the error. IMPORTANT: evaluateErrors needs - if asked - *combine* the calculation of the function value h(x) and the derivatives dh(x)/dx. This was a major performance issue. To do this, boost::optional<Matrix&> arguments are provided, and tin EvaluateErrors you just says something like
if (H) *H = Matrix_(3,6,....);
3) We are no longer using int or strings for nonlinear factors. Instead, the preferred key type is now Symbol, defined in Key.h. This is both fast and cool: you can construct it from an int, and cast it to a strong. It also does type checking: a Symbol<Pose3,'x'> will not match a Symbol<Pose2,'x'>
4) minor: take a look at LieConfig.h: it help you avoid writing a lot of code bu automatically creating configs for a certain type. See e.g. Pose3Config.h. A "double" LieConfig is on the way - Thanks Richard and Manohar !
1) eliminate methods no longer return a shared pointer. Shared pointers are good for Factors and Conditionals (which are also non-copyable), because these are often passed around under the hood. However, a BayesNet is simple a list of shared pointers and hence does not cost a lot to return as an object (which is compiler-optimized anyway: there is no copy). So, the signature of all eliminate methods changed to simply return a BayesNet<> object (not a shared pointer).
2) GaussianBayesNet::optimize is now replaced by optimize(GaussianBayesNet) and returns a VectorConfig and not a shared pointer
3) GaussianBayesNet and SymbolicBayesNet are now simply typedefs, not derived classes. This is desirable because the BayesTree class uses templated methods that return BayesNet<Conditional>, not a specific BayesNet derived class.
This simplifies Bayes nets quite a bit. Also created a Conditional base class, derived classes ConditionalGaussian and SymbolicConditional
Finally, some changes were needed because I moved some headers to .cpp
The main renaming that happened is
BayesChain -> BayesNet
ChordalBayesNet -> GaussianBayesNet == BayesNet<ConditionalGaussian>
SymbolicBayesChain -> SymbolicBayesNet == BayesNet<SymbolicConditional>
smallExample was changed to include two of the examples used in testConstrainedLinearFactorGraph
ConditionalGaussian was changed to make solve() virtual, as this is necessary for ConstrainedConditionalGaussian
''Limitations: ''
* Any given node can only have one constraint on it, but constraints can be of arbitrary size
* Constraints can only be specified as a blockwise system, where each block must be square and invertible to support arbitrary elimination orderings.
* ConstrainedNonlinearFactorGraph is disabled until a better solution for handling constraints in the nonlinear case is determined.
* Factors are now templated on the configuration type. Factor Graphs are now templated on the factor type and configuration type.
* LinearFactor is a factor on an FGConfig.
* LinearFactorGraph uses LinearFactor and FGConfig.
* NonLinearFactor is still templated on Config.
* NonLinearFactorGraph uses NonLinearFactors, but is still templated on Config.
* Tests and VSLAMFactor have been updated to reflect those changes.
(1) FactorGraph and NonlinearOptimizer now no longer have a .cpp file, but a -inl.h file as in [http://google-styleguide.googlecode.com/svn/trunk/cppguide.xml Google's C++ Style Guide]. This means if you expect to instantiate one of the functions in a cpp file, you have to include the -inl.h file.
(1) getOrdering is now in FactorGraph, and the non-linear version does *not* take a config anymore.
Long version: I made this change because colamd works on the graph structure alone, and should not depend on the type of graph. Instead, because getOrdering happened to implemented in LinearFactorGraph first, the non-linear version converted to a linear factor graph (at the cost of an unnecessary linearization), and then threw all that away to call colamd. To implement this in a key-neutral way (a hidden agenda), i had to modify the keys_ type to a list, so a lot of changes resulted from that.
Alex Cunningham
Frank Dellaert
Richard Roberts
Michael Kaess
Chris Beall