Calibration & abstention
The hard part isn’t finding the place. It’s knowing when you can’t.
Pic2Nav treats abstention as a first-class behavior. We measure coverage, precision when it answers, calibration error, and the cost of staying silent — because a wrong location is often worse than none at all.
Current snapshot
Measured from the live system on April 1, 2026
399
stored recognitions in the live system
88.73%
positive rate across feedback-bearing records
99
vectors in the deployed baseline index
14
unique places in the current place-grouped held-out split
Graded uncertainty
Not one answer or none. Five honest ways to respond.
The paper reframes abstention from a binary null into a graded interface. The product renders the same ladder — every result in the gallery and the scanner carries one of these states, not a bare coordinate.
Exact answer
Strong evidence supports a single point, returned with calibrated confidence.
Bounded guess
Useful but not exact — a region or coordinate with an uncertainty radius.
Evidence summary
Detected clues without a coordinate. The system explains what it sees.
Ask for more
Recoverable uncertainty — request GPS, a clearer sign, or a second angle.
Hard abstain
No location returned, with a reason tag. An honest no-clue case.
Research tracks
Three things we are actually working on.
The research page should describe the current system honestly: what is strong, what is experimental, and where the product and model stack currently meet.
Hybrid inference routing
The product is designed around evidence hierarchy: direct signals first, broader reasoning only when the image actually needs it.
Feedback-driven retrieval memory
Recognitions, confirmations, and corrections are persisted so the system can improve as an operational memory, not just a static model.
Backbone experiments
We test stronger vision backbones against the production CLIP baseline using deterministic place-grouped held-out evaluation.
System program
The stack is organized around evidence, not hype.
Direct evidence
EXIF GPS and visible-address shortcuts
The route checks deterministic signals first so the stack can return precise results without unnecessary model escalation.
Retrieval and priors
NaviSense V3
The ML service combines image embeddings, retrieval memory, and geospatial priors to narrow candidates before route-side validation.
Scene reasoning
Claude and Google Vision
When evidence is partial or messy, OCR, landmark hints, phone numbers, and scene-country reasoning become part of the decision path.
Why calibration
Confidence and correctness are not the same number.
Across 63 verified feedback items in the production snapshot, route reliability splits sharply. The most confident route is the least correct: navisense-ml reports a mean confidence above 1.0 while only 58% of feedback-bearing cases were right.
That gap is the whole argument for calibrated, method-aware confidence — and for abstaining when the evidence can’t carry the claim.
navisense-ml
Retrieval acceptance — the overconfidence case · 12 items
piexifjs-gps
EXIF GPS — direct metadata · 8 items
navisense-location-claude-name
Retrieval + name resolution · 3 items
claude-address-search
OCR + address resolution · 38 items
claude-immediate-uk
Regional scene reasoning · 4 items
Production snapshot · 2026-04-01 · feedback-bearing records only
Backbone evaluation
StreetCLIP is the first experiment that clearly moved the retrieval path.
We now have a deterministic place-grouped held-out evaluation from 38 canonical records spanning 14 unique places. That split is still small, but it is much more honest than the earlier tiny diagnostic slice.
The strongest current production path is still hybrid orchestration rather than a solved end-to-end geolocation regressor.
Claude-assisted address resolution currently outperforms direct ML acceptance in user-confirmed feedback-bearing production cases.
The held-out comparison is more honest than the earlier tiny diagnostic slice, but it is still directional rather than benchmark-grade proof.
Production baseline
CLIP ViT-B/32
The current deployed baseline is still the safer production default, but the direct geolocation head remains weak in absolute terms.
Experiment service
StreetCLIP
The first controlled backbone comparison shows a clear gain on the retrieval-driven path, even though the canonical corpus is still small.
Outputs
Publications, datasets, and the product surface stay connected.
Working paper
A systems paper framing Pic2Nav as a deployable hybrid photo geolocation stack.
Read publication notesDatasets and corpora
The project is moving toward stronger canonical evaluation and better place-grouped testing discipline.
Explore datasetsAPI and product surface
The research stack stays close to the actual interface rather than living separately from the product.
View API accessNext step
Test the research stack on real images.
The cleanest way to understand Pic2Nav is still to use it: upload an image, inspect the evidence path, and review how the system handles confidence and failure.
Current stance
Honest about what is solved, and what is not.
The research page now reflects the actual paper and live measurements: strong hybrid orchestration, improving retrieval, and a geolocation head that still needs better data and stronger evaluation.
Pic2Nav is best understood as a deployable geolocation stack for messy real-world photographs, not as a single solved benchmark model.