Building an accurate map of the road surface itself, not the buildings around it, matters for lane-level perception and high-definition maps in self-driving. Mesh-based reconstructions do this but tend to be costly to optimize and rough once the scene gets large.

This framework, called ROADGS-T in the text, represents the road as a grid of flat 2D Gaussian surfels, each carrying color, semantics, and geometry. Since roads are essentially thin sheets, that grid fits them better than full 3D Gaussian blobs or dense meshes, and it trims redundant, overlapping primitives during optimization. The grid also adapts, packing more surfels around lane markings, boundaries, and height changes while staying sparse over plain asphalt. A pose-refinement step leans on several nearby vehicle poses at once, weighting their height guidance by geometric consistency.

The abstract describes the design but stops short of numbers, so the paper is where its performance actually gets settled.