docs

Reference material — worked example

LUMEN-PS is a general shallow-surface scanner for mostly diffuse subjects. This folder uses a leaf because foliage motivated the project and makes fine relief easy to judge; the reconstruction is not limited to leaves. Paper, fabric, cardboard, bark, and other flat matte materials are also candidates. Avoid very shiny subjects, and use caution with rigid objects: raised, heavy, or sharp items can scratch or crack the scanner glass or damage its lid/mechanism.

A complete filtered run of the pipeline on a Kiwi leaf, captured on the HP LaserJet M1130/M1132 flatbed at 1200 dpi. These are compressed, downscaled images for documentation only — photos/diagnostics are JPEG, normal maps and anything with alpha are 8-bit PNG. The originals were the full-res lossless outputs; regenerate them by re-capturing and running the pipeline.

scans/cropped/

k0..k3.jpg — the four scans, rotated 90° between each (petiole steps left → down → right → up), cropped to a common scanner-space ROI around the leaf. 8-bit sRGB (this scanner's WIA driver caps at 8-bit). (Lossy previews here; a real run needs lossless PNG/TIFF captures — the CLI ignores JPEGs on purpose.)

results/

Final deliverables (downscaled to ≤1500 px):

• normal_gl.png / normal_dx.png — tangent-space normal maps
• albedo.jpg (linear — looks dark) / albedo_srgb.jpg — lighting-free base colour
• alpha.png — subject silhouette / opacity
• albedo_srgb_rgba.png, normal_gl_rgba.png — RGBA copies for a transparent plane
• height.jpg — Frankot–Chellappa integrated relief

These result previews include residual-guided bad-normal repair. The original full-resolution filtered sources are from sessions/kiwi-leaf-20260712-184906/out.

qa/

The diagnostic set (spec §8.4 — the re-render residual is the main trust signal):

• residual_scan0..3.jpg — |observed − re-rendered| per scan. Low and noise-like here → trustworthy solve. Final residual means ≈ 0.009 across all four scans.
• observed_scan* / predicted_scan* — the two sides of that comparison.
• mask_agreement.jpg — per-pixel count of valid samples (4/4 across the leaf).
• normal_preview.png — 8-bit preview of the normal map.
• subsurface_hint.jpg — R−B channel difference (vein/subsurface hint, §5.2).
• rejection_coverage.jpg — how often each pixel survived outlier rejection.
• misreg_repair.png — gray pixels were re-solved after rejecting one locally inconsistent observation; white pixels required selective inpainting.
• report.txt — fitted light vectors + residual stats for this run (self-calibrated: az0 = 90°, el = 45°).