Joint Multi-Condition Representation Modelling via Matrix Factorisation for Visual Place Recognition
The method jointly models places using multiple reference descriptors by stacking them and applying QR decomposition to create an orthonormal basis for each place. This training-free approach allows for projection-based matching, where a query is matched by minimizing the residual error after projecting it onto these learned subspaces. It enables robust localization across viewpoint and appearance variations.