Amplify motion

Breathing, a structure settling, the small sway of something that looks still. The motion pipeline makes these visible by amplifying how much the picture's detail shifts over time.

import vidmag

vidmag.magnify("baby.mp4", preset="motion", out="breathing.mp4")

Which filter to use

The motion pipelines differ only in how they pick out the frequencies to amplify, and the choice matters more than it looks.

Filter	Choose when	Cost
Two running averages (motion preset)	You want a broad band and the simplest behaviour, or you work frame by frame	Cheapest; needs only the previous frame
Butterworth	You want a defined band with a gentle edge, still frame by frame	Cheap; needs only the previous frame
Fourier	You want a sharply defined band and have the whole clip	Needs every frame at once

The motion preset uses the first — the authors' choice for the sleeping-child clip.

The two parameters that matter

alpha — how much to amplify. Raise it until the movement is clear, then stop. Past that, the picture tears into ripples and haloes: the method assumes the movement is small enough that shifting the image is nearly the same as scaling its detail, and that assumption fails first at edges.

lambda_c — the spatial cutoff, in pixels. Detail finer than this is amplified less, on purpose: fine detail is where the small-movement assumption breaks and where the sensor noise is. Raise it to reduce artefacts (and amplify less); lower it to do the reverse.

gentle = vidmag.magnify("clip.mp4", preset="motion", alpha=5, lambda_c=32)
strong = vidmag.magnify("clip.mp4", preset="motion", alpha=25, lambda_c=8)

Shimmering ripples mean too much alpha for the lambda_c in use. Raise lambda_c before lowering alpha.

Colour

chrom_attenuation scales how much the colour channels are amplified relative to brightness. The motion preset sets it to 0.1, so colour is amplified about a tenth as much. Colour noise is very visible and rarely carries the movement, so keeping it low is usually right.

A clip too long to fit in memory

Every pipeline here takes the whole clip at once. For something long, work in overlapping chunks and discard the overlap:

import numpy as np
import vidmag
from vidmag.io.video import load_video

video, info = load_video("long.mp4")
frames = np.clip(np.rint(video * 255), 0, 255).astype(np.uint8)

chunk, overlap, pieces = 300, 30, []
for start in range(0, len(frames), chunk):
    block = frames[max(0, start - overlap) : start + chunk]
    out = vidmag.magnify(block, preset="motion", fps=info.fps)
    pieces.append(out[overlap:] if start else out)

result = np.concatenate(pieces)

The overlap exists because the filters need history: without it each chunk restarts from nothing and the joins show.