Image Geolocalisation and Visual Place Recognition
Finding where a photo was taken, either on the globe or against a map of known places.
Tier Grades
Tier A: Frontier AI labs
feasible
Google trained PlaNet on millions of geotagged images.
Tier B: Indonesian AI labs
feasible
StreetCLIP is public and works zero-shot, so use needs little compute.
Tier C: Institution units such as ITB
feasible
StreetCLIP is public and works zero-shot, so a university lab can evaluate it on local photos.
Tier D: Organisations such as SEMAI (PoT, PoC, PoV)
stretch
Field photo verification is a possible use, but no industrial deployment is shown in the sources.
Where this comes from
Tiers: a) frontier AI labs, b) Indonesian AI labs, c) institution units such as ITB, d) organisations such as SEMAI delivering proof of technology, concept and value. Each rationale cites the anchors below.
Anchors
- ComputeStreetCLIP and PlaNet: training compute not stated in the abstracts.
- DataPlaNet: millions of geotagged images over thousands of geographic cells; StreetCLIP: zero-shot, beating supervised models trained on more than 4 million images.
- Team size (proxy)Proxy only (author count of an exemplar paper, not a headcount): 3 authors on the PlaNet paper (arXiv 1602.05314); 3 on StreetCLIP.
- Time horizonYears.
- Shelf-life signalsStreetCLIP reaches state of the art zero-shot by grounding CLIP, overtaking supervised classifiers such as PlaNet.
- Pipeline roleTier A trains global models; Tier B and Tier C evaluate them on Indonesian imagery; Tier D could use them to check field photo locations.
Where this comes from
Team size is a proxy (author counts of exemplar papers), not a headcount.
Who and What
Exemplars
Observed actors
Linked models
Links
R&D TopicImage Geolocalisation and Visual Place Recognition
Models
StreetCLIPStanford University, geospatial
PlaNet v2Google, geospatial
NetVLADINRIA, visionDELF (DEep Local Features)Google, vision+3 more
StreetCLIPStanford University, geospatialPlaNet v2Google, geospatialNetVLADINRIA, visionDELF (DEep Local Features)Google, vision+3 moreUse Casesnone recorded
Who Runs Itnone recorded
Models 7
HF-NetETH Zurich, robot policyproduced byDELF (DEep Local Features)Google, visionproduced by
MegaLocPolytechnic University of Turin, geospatialproduced byNetVLADINRIA, visionproduced byPlaNet v2Google, geospatialexemplarStreetCLIPStanford University, geospatialexemplarDKAZE visual localisation (TI)drivingproduced byPapers
- NetVLAD: CNN architecture for weakly supervised place recognitionRelja Arandjelović et al., arXiv preprint, 2015 founding
- MixVPR: Feature Mixing for Visual Place RecognitionAmar Ali-bey et al., WACV 2023 (per arXiv comment), 2023 state of the art
- GeoCLIP: Clip-Inspired Alignment between Locations and Images for Effective Worldwide Geo-localizationVicente Vivanco Cepeda et al., NeurIPS 2023 (per arXiv comment), 2023 state of the art
- AnyLoc: Towards Universal Visual Place RecognitionNikhil Keetha et al., ICRA 2024 (per arXiv comment), 2023 state of the art
- General Place Recognition Survey: Towards the Real-world Autonomy AgePeng Yin et al., arXiv preprint, 2022 survey
Where this comes from
The papers to read first: where the direction began, its current best, and a survey when one exists.
Sources
- arxiv.org/abs/1602.05314 https://arxiv.org/abs/1602.05314
- arxiv.org/abs/2302.00275 https://arxiv.org/abs/2302.00275
To cite this page: Project AURION, AI Use-Cases Research & Innovation, "Image Geolocalisation and Visual Place Recognition", https://aurion.semai.id/topics/image-geolocalisation-and-visual-place-recognition/, records of October 2026.