Modelpublic

aakanshajagga/lunarsync-geoprior-v1-1

sourceHugging Faceotherupdated 28d agoView on Hugging Face

0likes29downloads

Model Card

LunarSync GeoPrior v1.1

Cloud-trained, Hugging Face-compatible hybrid lunar image correspondence model for Chandrayaan-2 OHRC imagery.

Validated production mode

The supported mode composes source pixel → lunar ground coordinate → target pixel using PDS geometry metadata. A geometry-conditioned EfficientLoFTR-derived learned checkpoint is included for research and optional local refinement. The image-only fallback is disabled by default because it did not pass product-disjoint validation.

Evaluation

—Seven independent full-resolution OHRC acquisitions
—Five disjoint spatial checkerboard folds
—348,365 held-out control evaluations
—Degree-9 polynomial geometry prior, selected after a degree 3–9 sweep
—Median-of-fold-medians: 0.412 px
—Mean p95: 0.780 px
—Mean PCK@1 px: 99.773%
—PCK@2 px: 100%
—Learned parameters: 16,687,826

See evaluation_report.json for complete fold statistics, ablations, and the rejected image-only result.

Loading

python

from transformers import AutoConfig
config = AutoConfig.from_pretrained(
    "aakanshajagga/lunarsync-geoprior-v1-1",
    trust_remote_code=True,
)

Use geometry_prior.py and geoprior_coefficients.npz for the validated metadata-based correspondence path. Only supported Chandrayaan-2 OHRC product IDs listed in geoprior_products.json are covered by this release.

Limitations

This is not a general image-only lunar matcher. Predictions outside the supported acquisitions, without valid PDS geometry, or under unrelated sensors/resolutions are out of distribution. The learned fallback remains experimental.

Provenance

Training and evaluation were performed entirely on Kaggle cloud compute. Source imagery and geometry derive from Chandrayaan-2 OHRC products distributed under ISRO's applicable data policy.