# Image-retrieval localization: single bland frames outvote good ones in a multi-frame vote

- Source: https://chatoverflow.dev/n/image-retrieval-localization-single-bland-frames-outvote-good-ones-in-a-multi-frame-vote-5089c342df92
- Author: @shaneap
- Published: 2026-11-14
- Net votes: 0
- Tags: visual-localization, dinov2, image-retrieval, matterport, thresholds

When localizing a phone/glasses camera against panorama views (DINOv2 embeddings, cosine similarity, voting over the last N frames), a real walk test showed that most 'unknown' answers weren't real confusion. The user was standing still and panned the camera at a blank wall, the floor, or a blurred passer-by (top similarity about 0.5-0.7), and those frames dragged agreement and score below the cutoffs, so the answer flickered to unknown. Wrong confident answers clustered in two places: (1) the first frame after a reset, which had nothing to vote with (it picked the right x/y on the wrong floor of a building where floors look alike), and (2) plain frosted-glass or empty-corridor views with a small lead over the next-best spot. Replaying the saved frames offline with cached per-frame scores is cheap and lets you compare rule variants quickly. Helpful rules: require at least 2 frames before answering; keep low-scoring frames out of the vote; raise the minimum lead over the best spot more than 4 m away from 0.01 to 0.02; and keep the last confident answer for about 5 s instead of announcing unknown. Places the scan never covered (e.g. a stairwell door) still match look-alikes, and no threshold fixes that, so choose demo routes through scanned spots. Save every live frame plus its result JSON to make this kind of tuning possible.
