looming op• 資料種類:matrix × table → table
• 呼叫: import fullseye as fs; fs.ledger.fly_expansion_readout(flow, lattice, axis=None, weights=None)(要直接呼叫實作,import flyvision; flyvision.fly_expansion_readout(flow, lattice, axis=None, weights=None);從台帳取用則 opsflyvision.get("fly_expansion_readout"))
由整個眼的光流讀出碰撞時間——孔徑效應給出恰好 2 倍。
> 以下的詳細說明為原文 —— 摘要與標題已翻譯。
A fronto-parallel surface approached at speed v from distance d moves the line of sight at polar angle theta outward at sin theta cos theta / tau with tau = d/v. This op projects the measured flow onto the outward radial direction and fits that one number, 1/tau, by weighted least
squares::
1/tau = sum w f_r sin cos / sum w (sin cos)^2
and, independently, fits a linear field f = A x + b on the gnomonic plane and reads the divergence tr A, which sceneflow converts the same way (ttc = 2/div).
The reason to have both: a correlation-type detector only sees the motion component across a contrast edge. Over vertical stripes every edge is vertical, so the elevation component of the flow is lost (the aperture problem), the radial projection keeps on average half the expansion, and both readouts return 2 tau — the surface looks twice as far as it is. A checkerboard has edges in both directions and returns tau. That is the identity How et al. (2020) argued from, written as a number.
flow: (n, 2) azimuth/elevation flow per ommatidium (radians per
second), the return of fly_flow_from_directions.
lattice: the eye it was measured on (hexagonal or Vogel).
axis: the direction of travel = focus of expansion, body coordinates;
None = the lattice's optical axis.
weights: None or (n,) non-negative weights (e.g. 0 on the rim, where
fly_flow_from_directions reports nothing).
Returns a dict {"tau_s", "rate_1_s", "tau_div_s", "divergence_1_s", "aperture_ratio", "residual_rms", "n_ommatidia"} — the radial-fit time-to-contact and its rate, the divergence-fit time-to-contact and the divergence, the ratio of the smaller to the larger diagonal of A (1 for isotropic expansion, 0 under a pure aperture effect), the rms of the radial residual and how many ommatidia voted. A non-expanding field (rate <= 0) returns NaN for the time-to-contact — a documented non-finite, as in fly_tau_from_expansion: a receding surface has no time-to-contact.
Ground truth (pinned in the tests): the exact plane flow returns tau_s == tau and tau_div_s == tau to 1e-9 (the divergence is fitted on gnomonic velocities, where a plane approach is exactly linear); the same flow with its elevation component removed returns 2 tau from both readouts within 1% on a 10-degree field (aperture_ratio 0) — a hexagonal lattice is 6-fold symmetric, so sum cos^2 phi = sum sin^2 phi over it, and the residual is the sphere's departure from that plane.
Raises ValueError: a *flow* that is not (n, 2) for this lattice (an empty *flow* is refused), non-finite entries, a malformed *lattice* / *axis* / *weights*, all-zero weights, and fewer than 6 voting ommatidia.
• 範例資料目錄(下載 URL / 授權) —— 2-D 用 skimage.data(BSD/公有領域)加合成圖,3-D 給出真實資料源(Stanford/PDS 等)的下載 URL。
• 運算子來歷與參考文獻 —— 該運算子族所依據的研究/方法出處。
• 演算法的正典(作者、年份)與用途見上面的族使用指南。
• poc_zebra_stripes_fly_eye — py -3.11 examples/poc_zebra_stripes_fly_eye.py
table 作為輸入)fly_t4t5_field · fly_flow_from_directions · fly_matched_filter · fly_egomotion_from_flow · fly_eye_merge · fly_view_plane_texture · fly_hex_resample · lattice_moire_peak
looming)fly_lgmd_eta · fly_tau_from_expansion
*Provenance: flyvision.py — FLYVISION 運算子登記表。本條目由 tools/opdocs.py md 自動產生(請勿手動編輯)。*
© 2026 Kazufumi Furuse — Fullseye operator documentation. Licensed under Apache-2.0.