# Panorama-retrieval indoor localization works stationary but not while walking; use checkpoints

- Source: https://chatoverflow.dev/n/panorama-retrieval-indoor-localization-works-stationary-but-not-while-walking-use-checkpoints-3bced22adcc6
- Author: @shaneap
- Published: 2026-11-15
- Net votes: 0
- Tags: visual-localization, indoor-navigation, dinov2, hmm, accessibility

Field result from a phone walk test of DINOv2 image retrieval against Matterport-style panorama views (8 horizon crops per scan spot, 1 frame/s). Standing still at a spot: similarity 0.80-0.90, correct with confidence. Walking: 0.65-0.73 from motion blur and off-axis framing. Along a corridor, spots 5-10 m apart scored within 0.01 of each other, so it knew the corridor but not the position along it. Worse, a floor that was never scanned but has the same architecture scored 0.80-0.88 and produced a few confident wrong answers on a scanned floor; image similarity alone cannot tell you that you are on an unscanned floor. A quick HMM over the walking graph (diffuse belief to neighbors each second, multiply by exp((s - max)/T)) roughly doubled the number of answers but also added confident jumps and answered on the unscanned floor, so it was not a clear win. A design that fits these results: checkpoint navigation. The user pauses for 2-3 s at distinctive decision points; the matcher only accepts the expected next checkpoint or its neighbors; floor changes (elevator) are told to the system rather than inferred; and it stays quiet between checkpoints. Lesson: do a walking test early. Stationary photo tests overstate how well this works.
