Support gcc12 and ceres 2.1.0

[efernandez]

This MR has all the changes required to support gcc12 and Ceres 2.1.0.

$ g++ --version
g++ (GCC) 12.2.0

With Ceres 2.1.0 the ceres::LocalParameterization was marked as deprecated in favour of the new ceres::Manifold that implements the Plus and Minus methods; in Ceres 2.2.0 the ceres::LocalParameterization is already removed.

gcc12 spotted some issues:

• An out of bounds access in the CostFunctor defined for the Poin2DFixedLandmark unit test
• A maybe-unitialized warning in fuse_graphs related to boost::make_transform_iterator and a lambda that takes a reference from constraints or variables. I simply added -Wno-maybe-uninitialized for now
• A array-bounds warning due to this gcc12 bug, also mentioned in Eigen issue. I simply added -Wno-array-bounds, conditional on the compiler being gcc and >=12.0.

I also added SYSTEM to 3rd party headers to ignore warnings from those libraries.

The Minus Jacobian implementation of ceres::Manifold only computes the jacobian wrt the 1st argument, which differs from the fuse_core::LocalParameterization implementation, that computes it wrt the 2nd argument, which seems to be wrong. For that reason, I fixed the following bugs:

• Fixed the fuse_core::AutoDiffLocalParameterization
• Used ceres::AutoDiffManifold instead of fuse_core::AutoDiffLocalParameterization for Ceres >=2.1.0
• Updated the unit tests that had hard-coded expected Jacobians. They basically changed in signed. This only impacted the orientation 2D and 3D local parameterizations, which are the only ones provided.
• Fixed the analytic Minus Jacobians for the Orientation2DLocalParameterization and Orientation3DLocalParameterization, since they had the wrong sign. This bug was hiden by the bug in fuse_core::AutoDiffParameterization, which in a way was cancelling it out.

[efernandez]

Not sure why tests are failing now. I run all tests with both configurations, and all tests passed in both cases. That is, gcc 11.3.0 with Ceres 2.0 and gcc 12.2.0 with Ceres 2.1.

If you share the output errors with me, I can have a look.

[svwilliams]

Looks like some transient issues we are having with our own build server. I don't think it is related to this MR at all.

[jakemclaughlin6]

The general logic for the conditional compilation is as follows:

• If ceres is < 2.1, everything remains the same and LocalParameterization is used
• If ceres == 2.1, LocalParameterization is still supported for fuse::Variables, but everything in fuse switches to using Manifold, fuse::Variables implement a manifold from their existing LocalParameterization using the ManifoldAdapter
• If ceres > 2.1, LocalParameterization is completely removed, along with the ManifoldAdapter, everything must implement manifold directly

[clalancette]

It would be great to get this PR merged; the use of ceres::LocalParameterization is currently causing fuse_core to fail to build on Ubuntu 24.04 in ROS 2 Rolling: https://build.ros2.org/view/Rbin_uN64/job/Rbin_uN64__fuse_core__ubuntu_noble_amd64__binary/9/console

[svwilliams]

Yep, it's just not ready yet. I'm hoping to get some time next week to finish this up, and then port it over to ROS 2.

[svwilliams]

I've finished up the work on serializing the Manifolds. I didn't have permission to push to this fork/branch, so I created a new PR here for safe-keeping: #363

Once this is reviewed by my team, my intention is to merge this PR then rebase #363 on top of it. That should hopefully keep the contributions clear.

Let me know if you have any concerns on the changes that were made. And thanks for doing the vast majority of the work here. I'll be starting on a ROS 2 port of this work next week.

[jakemclaughlin6]

That works for me. The changes look fine to me, lets wait for @efernandez to see if its good with him as well

[svwilliams]

Reviews for the final set of changes is here: #363 Merging this PR and rebasing #363