Bench Results — 17-Sample Bone Depth Study

17 bone samples (“locations”), each photographed 4 times by a multi-LED rig. Depth was recovered per sample with Woodham photometric stereo (per-pixel surface normals from 3 single-LED images) + Frankot–Chellappa FFT integration.

Capture convention (verified on all 17 locations)

File	Lighting	Notes
`pN-1.jpg`	All LEDs on	Uniform reference; NOT used for the depth solve
`pN-2.jpg`	RIGHT LED	Light azimuth ~0°
`pN-3.jpg`	TOP/upper LED	Light azimuth ~40–90°
`pN-4.jpg`	LEFT LED	Light azimuth ~180°

Files in this folder

depth_summary — Google Sheet, one row per location (see columns below)
bone_depth_batch.py — the program that produced everything (re-runnable)
Image bundle (.zip) — 17× depth maps + 17× normal maps + 17× panel figures (drag the zip in, then right-click and extract, or just open it — the PNGs are inside)

The bundled PNGs — per-sample depth maps, normal maps, and panel figures — are not mirrored on this page.

depth_summary columns

Column	Meaning
`location`	Sample number 1..17
`width, height`	Working resolution used (640×360)
`n_lights`	LEDs used for the solve (3)
`azimuths_deg`	Light directions used (0; 90; 180 = right; top; left)
`depth_min/max`	Relative depth extremes
`depth_mean/std`	Mean is ~0 by construction; std = overall roughness
`relief_p2_p98`	Headline number: robust peak-to-valley surface relief (bigger = more pronounced 3-D structure on that sample)

What the numbers mean

Photometric stereo recovers SHAPE up to an unknown offset and scale, so depth is RELATIVE (working-pixel units), not absolute millimetres. The shape and the relative height differences are meaningful; the absolute value is not. To get true mm later you would add one coaxial point-light image and rescale — see the note at the bottom of bone_depth_batch.py.

How to re-run

pip install numpy opencv-python scipy matplotlib
# put the pN-M.jpg photos in a folder, set INPUT_FOLDER at the top of the
# script to point at it, then:
python bone_depth_batch.py

Outputs land in ./depth_outputs (the PNGs and the CSV in this folder).

Caveat

With 3 oblique, uncalibrated LEDs on a glossy specimen, fine relief is reliable but the very-low-frequency “overall tilt” of each map is approximate — inherent to photometric stereo, not a bug. Specular-glint removal is on by default.

Per-location depth estimates

One row per sample location. depth_px is the robust peak-to-valley relief in working-resolution pixels; depth_mm_at_30mm_FOV is the same relief converted to millimetres assuming a 30 mm field of view.

location	image_width_px	image_height_px	depth_px	depth_mm_at_30mm_FOV	max_relief_px	roughness_px
1	640	360	64.48	3.023	73.19	19.35
2	640	360	61.28	2.872	66.83	17.67
3	640	360	59.19	2.774	68.28	15.59
4	640	360	57.88	2.713	65.37	17.55
5	640	360	56.94	2.669	65.25	17.73
6	640	360	52.73	2.472	61.34	17.48
7	640	360	51.93	2.434	57.16	16.00
8	640	360	50.48	2.366	60.13	13.58
9	640	360	45.61	2.138	56.37	11.69
10	640	360	48.49	2.273	55.87	12.99
11	640	360	53.44	2.505	59.52	15.33
12	640	360	56.31	2.64	64.65	17.59
13	640	360	55.95	2.623	62.28	18.23
14	640	360	55.79	2.615	62.69	17.44
15	640	360	49.82	2.335	55.88	14.91
16	640	360	42.61	1.997	52.88	11.36
17	640	360	46.01	2.157	55.33	11.37

Across the 17 locations, mean peak-to-valley relief is roughly 2.5 mm (range 2.0–3.0 mm) under the 30 mm field-of-view assumption. These millimetre figures inherit the relative-scale caveat above: the pixel-to-mm conversion assumes the stated FOV rather than a metric calibration.