Downloads · 30 days
6
9% of all-time downloads
localfind-potsdam-onnx
ekremabiii/localfind-potsdam-onnx
localfind-potsdam-onnx is a machine learning model from ekremabiii. Use it for the machine learning task on the model card, and read the license before you ship it in a product. It is set up for onnx. The card lists the license as mit.
City-scoped visual place recognition for Potsdam, fully on-device. A photo never leaves the phone.
All-time downloads
67
Public
Repo size
679 MB
Likes
0
Public
Repo makeup
Click a slice to open those files.
.onnx679 MB · 100%
At a glance
- Library
- onnx
- License
- mit
- Access
- Public
- Created
- Jun 22, 2026
- Updated
- Jun 22, 2026
- SHA
- 8560b0a8
Model card
From the Hugging Face model README
LocalFind — Potsdam (on-device VPR)
City-scoped visual place recognition for Potsdam, fully on-device. A photo never leaves the phone.
Instead of global GPS regression (GeoCLIP), this uses retrieval: a frozen StreetCLIP ViT-L/14-336 encoder embeds the query, then nearest-neighbor against a gallery of 45,203 Mapillary street images (heading-aware deduped) returns the top-k coordinates. No training.
Pipeline
photo -> CLIP preprocess (336x336) -> [streetclip_vision_fp16.onnx] -> 768-d features
-> [potsdam_head.onnx] (L2-normalize -> cosine vs 45K gallery -> top-k -> coords)
-> top-k coordinates + similarity -> map + Street View verify
Accuracy (PoC)
- In-domain (street photo -> street gallery): ~19 m median top-1.
- Real arbitrary photos (Wikimedia mix incl. interiors/aerials): ~480 m median top-1 — bounded by viewpoint/content coverage of the street-only gallery. Best with on-street phone photos; use the top-5 + Street View verification.
Files
| File | Size | Role |
|---|---|---|
streetclip_vision_fp16.onnx | 609 MB | StreetCLIP vision encoder (336) |
potsdam_head.onnx | 70 MB | 45K Potsdam gallery + top-k |
Gallery imagery: Mapillary (CC-BY-SA). Built on StreetCLIP (geolocal/StreetCLIP).
- Library
- onnx
- License
- mit
- Created
- Jun 22, 2026
- Updated
- Jun 22, 2026