Introducing Street-Match
Manta's street-level arm. It takes a photograph from a rough region to the specific block, building or address, and says so when the evidence does not reach that far.
Geo-estimation gets you to the region. Street-Match gets you to the exact place. It is one of Manta’s two modes, alongside Geo-estimation: one model and one evidence chain, with Street-Match as the part that answers not roughly where on Earth, but this specific place.
It names a street or a building only when the evidence supports one. When it does not, it says so and returns the region instead. That is the rule the rest of Manta runs under: the precision it states is a limit on what it may claim, not a score printed next to a guess.
Why it matters
A region is useful for triage. It is not useful for verification. When an adjuster needs to confirm that a photo of storm damage was taken at the insured property, or a journalist needs to check that an image of a protest matches the claimed intersection, the answer has to be specific. “Somewhere in Miami-Dade” does not hold up. “The 400 block of NE 2nd Avenue” does.
Different cities share architecture, road markings and planting, and that is where a confident wrong street does the most damage. Street-Match is built for those lookalike cases: it commits to a place when the match holds, and keeps the uncertainty visible when it does not.
Benchmarks
The field calls this task visual place recognition: given a street-level photograph, find where it was taken. It has seven standard boards, scored by Recall@1, the share of photographs whose first answer is the right place.
| # | Model | Year | SPED | Pitts30k | MSLS-val | Nordland | AmsterTime | Tokyo 24/7 | SVOX | Mean |
|---|---|---|---|---|---|---|---|---|---|---|
| 1 | SAGE (8448-D) | 2026 | 98.9 | 95.8 | 94.5 | 96.0 | 83.5 | 97.5 | — | 94.37 |
| 2 | Street-Match 1.0 ◇ | 2026 | 95.39 | 94.95 | 94.60 | 97.21 | 74.25 | 98.43 | 98.74 | 93.37 |
| 3 | SAGE-L (no InteractHead) | 2026 | 92.1 | 94.7 | 94.2 | 94.8 | 65.6 | 98.1 | 98.8 | 91.19 |
| 4 | BoQ | 2024 | 92.5 | 93.7 | 93.8 | 90.6 | 63.0 | 98.1 | 99.0 | 90.10 |
| 5 | SuperPlace NVL-FT² (B) ◇ | 2025 | 87.5 | 93.7 | 94.3 | 91.4 | 62.3 | 96.8 | 98.6 | 89.23 |
| 6 | EMVP ◇ | 2024 | 94.6 | 94.0 | 93.9 | 88.7 | 65.6 | 96.8 | — | 88.93 |
| 7 | SuperVLAD | 2024 | 93.2 | 95.0 | 92.2 | 91.0 | 63.9 | 95.6 | — | 88.48 |
| 8 | SALAD | 2024 | 92.1 | 92.4 | 92.2 | 90.0 | 58.8 | 94.6 | 98.2 | 88.33 |
| 9 | FoL | 2025 | 92.1 | 93.9 | 93.1 | 87.8 | 64.6 | 96.2 | — | 87.95 |
| 10 | SALAD-CM | 2024 | 89.5 | 92.6 | 94.2 | 95.6 | 57.8 | 96.8 | — | 87.75 |
| 11 | SelaVPR † | 2024 | 88.6 | 92.8 | 90.8 | 87.3 | 55.2 | 94.0 | 97.2 | 86.56 |
| Not ranked: reports fewer than 6 of 7 benchmarks | ||||||||||
| — | CricaVPR | 2024 | 91.3 | 94.9 | 90.0 | 90.7 | — | 93.0 | — | 91.98 |
| — | MixVPR | 2023 | 84.7 | 91.5 | 88.0 | 76.2 | — | 85.1 | — | 85.10 |
| — | EigenPlaces | 2023 | 70.2 | 92.5 | 89.1 | 71.2 | — | 93.0 | — | 83.20 |
| — | CosPlace | 2022 | 75.5 | 88.4 | 82.8 | 58.5 | — | 87.3 | — | 78.50 |
| — | NetVLAD | 2016 | 70.2 | 81.9 | 53.1 | 6.4 | — | 60.6 | — | 54.44 |
| — | SuperPlace NVL-FT² (L) ◇ | 2025 | — | 94.1 | 94.5 | — | — | 97.1 | — | 95.23 |
| — | EffoVPR † ◇ | 2025 | — | 93.9 | 92.8 | — | — | 98.7 | — | 95.13 |
Seven boards in, Street-Match 1.0 ranks second of the eleven models that report at least six. The model above it is SAGE: peer reviewed at ICLR 2026, weights released, and ahead on five of the seven boards. Second is what the numbers say, so second is what we say.
Nordland is the clear win: 97.21 against SAGE’s 96.0 over 27,592 photographs, a lead six times its own noise floor on the largest split in the table. MSLS-val, Tokyo 24/7 and SVOX are ties; their splits are small enough that a few tenths of a point decide nothing. Pitts30k and SPED are losses, by 0.85 and 3.51.
AmsterTime is the one that matters most. It is the cross-era task, matching an archival photograph to the street as it stands today, and it is the work the product does every day. At 74.25 Street-Match is clear of every other model in the table, and 9.25 behind SAGE. The board we care about most is the one we trail by the widest margin, and it is where the work is going. The full method, the noise floors and every citation are in the Manta post.
Availability
Street-Match is available now and runs inside Manta, which is available to organizations only, inside a reviewed workflow. It is not on the consumer web app or a self-serve API key. Coverage is expanding city by city, starting where it is deepest. We cannot hand out weights for anyone to rerun our rows, so we offer the next best thing: send us the hardest images you have, and we will send back what it produces.
Go deeper: Introducing Manta · Newsroom case study