lamina op• Data kinds: matrix → matrix
• Call: import fullseye as fs; fs.ledger.fly_lamina_filter(movie, dt_s, tau_adapt_s=0.2, tau_lp_s=0.02, mode='divisive', floor=0.001) (to call the implementation directly, import flyvision; flyvision.fly_lamina_filter(movie, dt_s, tau_adapt_s=0.2, tau_lp_s=0.02, mode='divisive', floor=0.001); from the registry, opsflyvision.get("fly_lamina_filter"))
Photoreceptor adaptation + the lamina's band-pass: intensities in, contrast out.
The first thing the optic lobe does to a picture is throw away its brightness.
A photoreceptor adapts to the running mean light level and the large monopolar
cells (L1/L2) report the *deviation* from it, so the same scene at dawn and at
noon arrives at the motion detectors as the same signal. Two closed-form
stages, in that order:
1. adaptation — a first-order low-pass of time constant *tau_adapt_s*
per ommatidium is the adaptation state `a(t). mode="divisive"`
returns the Weber contrast `(x - a)/(a + eps) (`eps = floor *
mean(x)``, so it scales with the picture and a dark ommatidium cannot
divide by zero); `mode="subtractive" returns x - a`, which is the
same high-pass without the gain control.
2. membrane — a first-order low-pass of time constant *tau_lp_s*, the
cell's own bandwidth.
movie: `(T, n) intensities, rows = time. mode="divisive"` refuses a
negative entry (a negative light level is not a measurement) and an all-zero
movie (its contrast is 0/0, which would be fabricated rather than measured).
dt_s: sample interval, seconds. tau_adapt_s / tau_lp_s: the two time
constants, seconds. floor: the divisive guard, relative to the mean intensity.
Returns `(T, n)` float64 contrast.
Ground truth, both exact rather than approximate:
• Weber invariance. In `"divisive"` mode, scaling the whole movie by
any positive constant returns *the same array* — both `a and eps`
scale with it. That is the point of the stage and the tests pin it to
machine precision.
• The transfer is the product of the two first-order filters. With
`A = 1 - exp(-dt/tau_adapt) and B = 1 - exp(-dt/tau_lp)`, the
steady-state gain at angular frequency `w` is
`|1 - H_A(w)| * |H_B(w)| where H(w) = C/(1 - (1-C) exp(-i w dt))` —
a band-pass that blocks DC exactly and is measured at four frequencies in
the tests.
Raises `ValueError`: a non-2-D / too-short / non-finite *movie*, a movie
over :data:MAX_MOVIE_ELEMENTS, a non-positive *dt_s* / *tau_adapt_s* /
*tau_lp_s*, a non-positive *floor*, an unknown *mode*, and (divisive only) a
negative or all-zero movie.
• Sample-data catalog (download URLs / licences) — 2-D uses skimage.data (BSD/public domain) plus synthetic images; 3-D lists download URLs for real data sources (Stanford, PDS, …).
• Operator provenance and references — the sources of the research/methods this op family came from.
• The canonical algorithm (author, year) and its uses are named in the family usage guide above.
• poc_fly_optomotor_steering — py -3.11 examples/poc_fly_optomotor_steering.py
matrix as input)fly_onoff_split · fly_t4t5_field · fly_flow_from_directions · fly_egomotion_from_flow · fly_hs_readout
lamina)*Provenance: flyvision.py — FLYVISION operator registry. This per-op note is generated by tools/opdocs.py md (do not hand-edit).*
© 2026 Kazufumi Furuse — Fullseye operator documentation. Licensed under Apache-2.0.