chiikabu-labs/chipoint-v2-geolocation

Chipoint v2 — Image Geolocation

Predict the geographic location of a photo using the full chiikabu-labs/chipoint-2 model — both arms.

How it works

1. 1.Encode: the input image runs through three frozen towers — SigLIP2 SO400M-256 and GOPT ViT-g/16 (both open_clip webli checkpoints) and DINOv3 ViT-L/16 (loaded from the ungated ModelScope mirror of facebook/dinov3-vitl16-pretrain-lvd1689m) — each projected by its trained MLP head to a unit-norm 512-d vector.
2. 2.Route: a router network reads the DINOv3-projected query standardized by the checkpoint's μ/σ;
3. 3.Retrieve (street arm): every enabled tower scores cosine similarity against its own 4.89M-row OSV-5M gallery (street_gallery/proj_head_gallery_512_{dinov3,gopt,so400m256}.npy); per-tower z-scored similarities are summed over the union of top-200 lists and the best 64 candidates kept.
4. 4.Retrieve (photo arm): one unified 14,541,898-row projected photo gallery (photo_gallery/emb_*.npy). A startup self-check identified that it lives in the DINOv3 projection space, so queries use the DINOv3 tower; all four shards are scanned exactly like the street galleries.
5. 5.Predict (cluster-consensus fallback): a global weighted mean over candidates spanning multiple continents is meaningless, so a cluster is seeded at the top-1 fused candidate and the inverse-sim⁴ similarity-weighted spherical mean is taken over that cluster only — candidates within 250 km great-circle of the seed, progressively relaxed to 500 / 1000 km when fewer than 3 candidates fall inside. The raw top-1 candidate is also shown alongside.

Galleries stay on disk as numpy memory maps and are streamed through the GPU in 250k-row chunks, so VRAM use stays flat no matter how large the index grows.

Reranker status

The chipoint-2 rerankers (streetview_k512_joint.pt, photoarm_k768_thr.pt — SR/ThrNet heads over a 23-feature stack) are wired in app.py but currently disabled (ENABLE_RERANKER = False). The exact training-time formulas for several features were never published (rank/agreement semantics; the merged-head features reference gallery_merged_cells.npz, which is row-aligned to the photo gallery only), so this demo serves the honest fused-retrieval fallback rather than guessing feature values.

License

The model is licensed under MIT. See the model card for details.