LOCATION FINDERv0.9 prototype

Technical note

The match engine

Location Finder scores candidates image-to-image. Your reference photo is compared against a Street View frame captured for each candidate in your search area, and the structured variables you set — search area, location type, architectural style, dominant colour or material — act as weights and filters on top of that visual similarity.

Listing and place metadata carries no reliable description of architectural style or facade colour, so it is never used as the primary signal. The whole pipeline runs inside this application, which keeps image work, scoring and quota enforcement in one place and avoids the extra latency of an outside automation service.

Pipeline

  1. Stage 1

    Brief intake

    Reference image plus variables are validated and geocoded into a bounded search area.

  2. Stage 2

    Candidate harvest

    Places search over the bounded area, plus password-free public MLS listings for the same area.

  3. Stage 3

    Frame capture

    One Street View frame per candidate at the best-facing heading, cached so a repeat search costs nothing.

  4. Stage 4

    Vision pass

    Reference and candidate frames are embedded once; similarity is cosine distance, not a text comparison.

  5. Stage 5

    Weighted score

    Visual similarity and structured variables combine into a single 0-100 match, filtered by your threshold.

  6. Stage 6

    Return

    Address, lat/long, municipality, match %, Street View frame and map pin — plus CSV export.

Scoring weights

45%

Reference image similarity

Vision embedding distance between the uploaded photo and the Street View frame.

18%

Location type

Expanded categories including accommodation, agriculture, airports, amusement parks, bridges, businesses, government, hospitals, lakes, roads, schools, sports, film studios, warehouses, and more.

15%

Architectural style

Style label inferred from the candidate image, not from listing text.

10%

Dominant colour / material

Colour histogram plus material classification of the facade region.

12%

Search area

Country / region / municipality agreement, resolved through geocoding.

Unspecified variables are treated as neutral, so a thin brief still returns ranked results instead of an empty set.

Cost and risk controls

  • Street View is billed per image. Frames and embeddings are cached per coordinate, so the second search over a neighbourhood is near-free and a bounded area caps the worst-case spend of any one search.
  • Search quotas are enforced server-side against the plan, not in the browser, so image spend cannot outrun revenue.
  • MLS coverage varies by region. Only password-free public listings are read, and each result is labelled with its source so nothing appears more authoritative than it is.
  • Vision confidence is shown, not hidden. Every result exposes its per-variable breakdown so a scout can see whether a 78% came from the image or from geography.

Not wired in this build

Google Maps, Street View Static, Places and Vision keys, MLS feed readers, and billing are all absent by design in this prototype. Everything on the console runs against sample data so the interaction model and scoring can be reviewed before any metered API is switched on.