Modelpublic

aakanshajagga/lunarsync-hybrid-v2

sourceHugging Faceotherupdated 28d agoView on Hugging Face

0likes31downloads

Model Card

LunarSync Hybrid v2.0

Shipping, cloud-trained correspondence pipeline for seven overlapping Chandrayaan-2 OHRC acquisitions.

Validated result

Metric	Result
Five-fold held-out controls	342,408
Median error	0.412 px
p95 error	0.770 px
PCK@1 px	99.812%
PCK@2 px	100.000%
Held-out acquisition PCK@1	99.776%
Visual gate test AUROC	0.962
Visual gate false-accept rate	0.75%
10k-point CPU latency	114.0 ms

The degree-9 geometry maps are refit independently in every spatial fold. Visual train, validation and test acquisitions are product-disjoint.

Modes

—metadata_geometry (production): deterministic pixel→lunar-ground→pixel correspondence.
—visual_gate (production safety): applies a local visual shift only above the validation-calibrated threshold; otherwise retains geometry.
—image-only learned global matching is not claimed. Failed refiners are documented in evaluation_report_v2.json.

Use

python

from huggingface_hub import snapshot_download
from inference import LunarSyncShippingPipeline
root=snapshot_download('aakanshajagga/lunarsync-hybrid-v2')
pipe=LunarSyncShippingPipeline(root)
result=pipe.correspond(source_product,target_product,[[1000,1000]])

Scope and limitations

This release supports the product IDs in geoprior_products.json. It requires PDS geometry metadata and does not generalize to arbitrary lunar imagery or TMC-2 without new calibrated data.

Free cloud demo

Open the browser-side Hugging Face Static Space. Geometry inference and the safety gate run client-side; no paid runtime is required.