dtof op• Datenarten: histcube → depth
• Aufruf: import photoncount; photoncount.dtof_cube_depth(cube, bin_ps=100.0, mode='peak', offset_ps=0.0, min_counts=1.0, empty_value=0.0, subtract_background=False) (oder opsphoton.get("dtof_cube_depth"))
Tiefenkarte aus einem `(H, W, T)`-Photonenhistogrammwürfel — die dToF-Inversion.
> Die ausführliche Beschreibung unten ist der Originaltext — Zusammenfassung und Überschriften sind übersetzt.
The array version of :func:dtof_depth, with the same four *mode* estimators
and the same `t_flight = t_measured - offset_ps` sign convention. The time
axis is last: a `(D, H, W)` voxel volume passed in here would be read as
`W` time bins and return a plausible-wrong depth map, so the shape is
checked and the error message says exactly that.
Pixels whose total counts are below *min_counts* — and, in the peak-based
modes, pixels whose histogram is exactly flat (no peak to find, so
`argmax` would report bin 0 for every one of them) — are set to
*empty_value* (default 0.0, a value no real return can have since
`d > 0`). Set
`empty_value=float('nan')` if you would rather propagate a NaN — that is an
opt-in, never the default, because a NaN depth map silently poisons every
downstream reduction.
Where a sub-bin *mode* cannot be applied to a pixel — the peak is in the
first or last bin, or the three samples are flat / non-positive for the log
fit — that pixel **falls back to the bin-centre (`"peak"`) estimate**. A
per-pixel exception would be useless on a megapixel cube; the fallback is
documented here and pinned in the tests, and it degrades to the coarser
estimator rather than to a wrong one.
Ground truth: on a noiseless simulated cube of a tilted plane from 1.0 to
3.0 m (32x32 pixels, 256 bins x 100 ps, 500 ps IRF) the RMS depth error is
4.39 mm for `"peak", 3.2e-16 m for "centroid"`, 0.114 mm for
`"parabolic" and 1.6e-8 m for "gaussian"`. With Poisson noise (20
signal + 5 ambient photons per pixel, seed 0) the same four give 19.9 mm,
164.8 mm (background subtracted), 18.7 mm and 19.2 mm — at 20 photons the
estimator choice is worth about 6%, and the centroid is 8x worse than doing
nothing clever at all.
Returns a float64 `(H, W)` depth map in metres.
Raises `ValueError`: a cube that is not 3-D / has fewer than 2 time
bins / holds negative counts / exceeds :data:MAX_CUBE_ELEMENTS, a
non-positive *bin_ps*, an unknown *mode*, a negative *min_counts*, a
non-finite *empty_value* other than NaN, and — instead of returning negative
distances — an *offset_ps* that exceeds the measured arrival time of any
valid pixel (a mis-signed or mis-scaled calibration delay).
• Leitfaden zur Familie photon_timeresolved
• Katalog der Beispieldaten (Download-URLs / Lizenzen) — 2-D nutzt skimage.data (BSD/Public Domain) plus synthetische Bilder, 3-D nennt Download-URLs echter Datenquellen (Stanford, PDS, …).
• Herkunft und Literatur der Operatoren — die Quellen der Forschung/Verfahren, auf denen diese Operatorfamilie beruht.
• Der kanonische Algorithmus (Autor, Jahr) und seine Anwendungen stehen im Familienleitfaden oben.
• photon_timeresolved — py -3.11 examples/photon_timeresolved.py
depth als Eingabe)dtof)dtof_depth · dtof_cube_simulate
*Provenance: photoncount.py — PHOTON Operator-Registry. Diese Notiz wird von tools/opdocs.py md erzeugt (nicht von Hand bearbeiten).*
© 2026 Kazufumi Furuse — Fullseye operator documentation. Licensed under Apache-2.0.