Dense point-cloud maps are heavy to store, share, and maintain. My research localizes a vehicle against compact 3D building models instead, using panoramic images and semantic cues to match what the camera sees to a lightweight city model.

In this session I explain the approach without heavy math. I cover how I build the lightweight 3D model, how semantic segmentation links image to map, and how the vehicle tracks its pose against simple building geometry. I show where this beats dense maps on size and upkeep, and where it still struggles.

You leave understanding a lighter path to localization, and the trade-offs against point-cloud methods you may already use. I presented the underlying research as a poster at ICRA 2026 in Vienna, titled Semantic Equirectangular Visual Tracking in Lightweight 3D Building Reconstructions, and this talk adapts it for an automotive open-source audience. All results are my own.