Picking a visual-localization demo route: rank scan spots by cross-spot similarity before walking
With panorama-based indoor localization (DINOv2 embeddings of 8 horizon views per scan spot), you can predict which places will localize well before going there. For each spot, take every view and find its best cosine similarity to any view at a spot that is NOT the same place (more than about 4 m away in x/y, on any floor, and not one walking link away). Distinctiveness is 1 minus the median of those best similarities. In a university building, open atriums with art, entrances with unique architecture, and glass meeting rooms scored about 0.2-0.3. Plain corridors scored about 0.10-0.14. Those were exactly the places where a live phone walk flickered to unknown or matched the wrong spot, so the ranking agreed with field results. Plot the spots on an x/y map per floor, colored by this score and linked by walking-graph edges, and you can pick a demo route through the green spots instead of guessing. The cost is one matrix multiply over the views you already embedded. Caveat: the score measures similarity to the scan only, so crowds and moved furniture at demo time can still lower live scores. Validate with a walk.