You can deploy this model and start earning money today!

ekremabiii/
localfind-geoclip-onnx

Name: localfind-geoclip-onnx
Availability: InStock

Views

No views yet

LocalFind GeoCLIP ONNX

On-device photo geolocation models for the LocalFind app. A photo never leaves the device — inference runs fully locally via ONNX Runtime (CoreML on iOS, NNAPI/XNNPACK on Android).

This is an ONNX port of GeoCLIP, split into two stages so the heavy gallery matmul stays inside the graph.

Pipeline

photo
  -> CLIP preprocess (224x224, see config.json)
  -> clip_vision_unnorm_fp16.onnx   (pixel_values -> 768-d, UN-normalized, fp16)
  -> head.onnx                      (MLP 768->512 -> normalize -> 100K gallery matmul -> top-k)
  -> topk_coords [20,2], topk_probs [20]
  -> cluster-based confidence (haversine spread of the top-k)

The vision stage is openai/clip-vit-large-patch14 get_image_features (un-normalized, pre-visual_projection-normalize tensor) so it stays a drop-in for the GeoCLIP MLP head. The head bakes in the precomputed 100K GPS gallery location embeddings (fp16 storage, fp32 matmul) and the gallery coordinates.

Files

File	Size	Role
`clip_vision_unnorm_fp16.onnx`	608 MB	CLIP ViT-L/14 vision encoder, fp16
`head.onnx`	107 MB	GeoCLIP MLP + gallery + top-k
`config.json`	—	I/O names, dtypes, preprocessing constants, sha256

Accuracy

Verified bit-exact against the reference PyTorch GeoCLIP pipeline. On the Eiffel Tower test image the full ONNX pipeline lands the top-1 within 0.05 km of ground truth (max top-20 coordinate deviation vs reference: 0.048 km).

License

MIT. Built on GeoCLIP (MIT) and CLIP (OpenAI).

ekremabiii/localfind-geoclip-onnx