Towards a Hybrid Scene Representation in VSLAM

In their basic form, Visual Simultaneous Localization and Mapping (VSLAM) systems rely on metric maps to produce meaningful camera pose measurements. As researchers shift their focus towards integrating higher levels of data representations, we argue that the performance of such systems requires a mutually beneficial, tight integration between metric, topological and semantic maps. Therefore, we emphasize the need for a hybrid locally semi-dense and globally sparse metric map and explore the multi-level hybrid map representation interactions for VSLAM and how they can be mutually beneficial.