synthesis op• 資料種類:無 → video(僅由參數決定的運算子 —— 不接受影像或資料輸入)
• 呼叫:import motionmag; motionmag.synthesize_translation(shape=(64, 64), frames: 'int' = 32, amplitude_px=0.5, frequency_hz=4.0, fps=32.0, direction_deg=0.0, wavelength_px=(8.0, 16.0), contrast=0.4, offset=0.5, noise_sigma=0.0, seed: 'int' = 0) -> 'np.ndarray'(或 opsmotionmag.get("synthesize_translation"))
位移有閉合解的片段 -> `(T, H, W)` 影片。
> 以下的詳細說明為原文 —— 摘要與標題已翻譯。
The scene is a stationary two-axis sinusoidal grating that is translated,
frame by frame, by
`d(t) = amplitude_px * sin(2*pi * frequency_hz * t / fps)`
along *direction_deg* (0 deg = towards +x / increasing column). The
translation is applied as a Fourier phase ramp, which for a pattern that is
periodic on the grid is the *exact* band-limited shift — no interpolation
kernel, no resampling error, so `d(t)` is ground truth to machine
precision and sub-pixel amplitudes are meaningful.
`wavelength_px is either one number (both axes) or `(lambda_x,
lambda_y)``. **The default deliberately makes the two axes different
octaves**, and that is not cosmetic: if the horizontal and vertical gratings
share a radial frequency they land in the *same* sub-band, whose local phase
is then the phase of a sum of two moving components rather than of one. The
phase of a sum is not linear in the displacement, so scaling it does not
scale the motion — measured, a 64x64 clip built with a single wavelength on
both axes magnified at `alpha = 2 came out at 0.939 * (2 d)` instead of
`2 d, a 6.1 % error that does *not* shrink as d` shrinks. Separating
the octaves puts one component in each band and the relation becomes exact.
This is the standard narrow-band condition of phase-based processing, made
visible in the synthetic instead of hidden.
Each wavelength is snapped so that a whole number of cycles fits the frame
(`cycles = max(1, round(W / wavelength_px))`, effective wavelength
`W / cycles`); without that the pattern is not periodic on the grid and the
Fourier shift would wrap a discontinuity across the border. With the default
64x64 frame and `(8, 16)` the snap is exact (8 and 4 cycles).
Values are not clipped into `[0, 1]`: clipping is a nonlinearity that
would break the exact translation this function exists to provide. With the
defaults the samples lie in `[offset - contrast, offset + contrast]`.
*noise_sigma* adds zero-mean Gaussian sensor noise (after translation, drawn
from `numpy.random.default_rng(seed)`) — the term that makes an SNR
measurable at all.
This is the counterpart of `photoncount.tcspc_simulate`: a forward model
good enough to close the loop on the analysis operators in this module.
• 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。
• motion_magnification — py -3.11 examples/motion_magnification.py
• poc_motion_magnification — py -3.11 examples/poc_motion_magnification.py
• quaternion_monogenic — py -3.11 examples/quaternion_monogenic.py
video 作為輸入)temporal_bandpass · temporal_band_power · band_snr · motion_magnify · phase_displacement · displacement_series
synthesis)—
*Provenance: motionmag.py — MOTIONMAG 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.