With default settings (no filter), slow drift swamps the audio: 0.3 px of drift cuts in-band level by 12 dB

[joeljose]

Severity: High. The README tip says "Start with default settings and adjust from there", but the defaults fail on realistic footage.

Problem

Without -fl/-fh, postprocess_phase_signals applies no filtering. The phase signals are measured relative to frame 0, so any slow motion (camera or tripod creep, thermal drift, the object settling) appears as a large low-frequency ramp. After sub-bands are added, the output is min/max normalised to [−1, 1], so the ramp takes up almost all of the range and the audio shrinks to a small fraction of it. Any DC offset also affects the cross-correlation alignment.

Evidence (synthetic, 0.05 px, 440/660/250 Hz tones, 2200 fps, 0.5 s)

added linear drift over the clip	default: correlation with truth	default: in-band (100–1000 Hz) RMS	-fl 100: correlation	-fl 100: in-band RMS
0 px	1.000	0.504	0.997	0.485
0.3 px	0.264	0.127 (−12 dB)	0.998	0.504
3 px	−0.025	0.016 (−30 dB)	0.952	0.426

A drift of 0.3 px over half a second is small for real footage, yet it makes the default output nearly useless.

Suggested fix

• Make a gentle high-pass the default, e.g. -fl defaulting to max(20 Hz, fps/200). Keep --no-filter for raw analysis.
• Always remove the mean and linear trend of each band before alignment and summing.
• Normalise robustly: scale by a high percentile (e.g. the 99.9th of |x|) or to a target RMS, and clip, instead of min/max, which a single transient (a knock or a dropped frame) can dominate.
• Update the README tip.

Acceptance criteria

• Default settings on the synthetic 0.3 px drift case give correlation ≥ 0.95.
• A test covers drift rejection with the default settings.