Pipeline: Render-and-Match
01 PREPAREPDS4 crop, north-up from the geometry grid, common GSD
02 RENDERDEM shaded by Hapke at the source Sun
03 MATCHRoMa v2 (DINOv3) dense warp; SIFT baseline
04 GEOMETRY8×8 grid-balanced sampling → MAGSAC++
05 VERIFYRMSE · inliers · coverage entropy → verdict
06 OUTPUTGeoTIFF · match CSV · metrics JSON
Registration runs
Benchmark (Apollo 11, known ground truth)
Source images are warped by a known transform, so the error is exact: GT RMSE = RMS error of the estimated transform over a 32×32 grid, in source pixels. "Sub-px success" = not rejected and GT RMSE < 1 px, over 3 crops.
| loading… |
Register your own pair
Chandrayaan-2 PDS4 products: cut them first with scripts/prep_ch2.py, which writes src.tif, ref.tif and dem.tif.
Read these results honestly
- Real Chandrayaan-2 runs have no independent check points, so RMSE there is transform self-consistency on the inliers, not ground-truth error.
- Synthetic (SYN) sources come from the same renderer CHITRA uses, so they are an upper bound for Render-and-Match.
- "Sub-pixel on OHRC↔NAC" is a target that these runs support. Literature numbers (SIFT IIRS→WAC ≈73 m, a preprint; CNSFM 72.3 % polar) are cited baselines, not ours.