Extending ROSETTA beyond outdoor driving environments and rigid rotational alignment

Extend the ROSETTA training-free local alignment framework to indoor and aerial environments, develop nonlinear alignment models beyond the special orthogonal group SO(K), and construct incremental variants that update the alignment estimate online.

Background

ROSETTA estimates a spatially varying feature-space rotation in SO(K) between independently trained, frozen encoders. The experiments evaluate outdoor driving datasets and use a per-window rotational model with sequential estimation, leaving broader environments, richer deformation models, and online updating outside the demonstrated scope.

The conclusion explicitly identifies three unresolved directions: applying the method to indoor and aerial settings, replacing the SO(K) assumption with nonlinear alignment models, and developing incremental estimators capable of updating alignment estimates online. These directions concern extending the method beyond its current environmental, geometric, and operational limitations.