Eve library

Motion and temporal quality (2026-09-10)

Guide

Revision 1

On this page

Evidence for 182.C.19.06: motion smoothness, temporal consistency, optical flow outliers, frame duplication, freeze, judder, warping, and background instability.

Method#

Seven 16-frame sequences built from the committed frame ../2026-09-10-openrouter-face-detection/two-foreground-subjects-plus-bystanders.png by real geometry — a moving crop window, a copied buffer, a remap — so each sequence genuinely has the property it is named for. Analysed with dense Farnebäck optical flow at OpenCV's documented defaults.

Results#

Sequence identical pairs longest run mean flow flow SD max accel local incoherence warp-measurable pairs
smooth-pan (control) 0 0 18.28 0.94 1.03 1.170 0
static (control) 15 15 0.00 0.00 0.00 0.0004 15
duplicated-frame 1 1 16.70 4.80 18.31 1.121 1
freeze 5 5 11.77 8.58 18.48 0.802 5
judder 9 2 7.07 8.68 18.48 0.472 9
warping 0 0 18.28 0.92 1.00 1.215 0
background-instability 0 0 3.99 1.83 4.83 0.348 0

Region flow (subject box against the rest of the frame):

Sequence subject flow background flow ratio
background-instability 0.420 4.781 11.4
smooth-pan 20.203 17.859 0.88

Two findings where the obvious approach points the wrong way#

Judder has MORE identical frames than a freeze#

The judder sequence has 9 identical frame pairs against the freeze's 5, because 3:2 pulldown repeats frames all the way through while a freeze holds one. A detector counting identical pairs would call judder the worse freeze.

What separates them is the longest consecutive run: 2 for judder, 5 for the freeze. Distributed repeats are a frame-rate conversion; a consecutive run is a stall.

And a freeze is not a static shot, which has an even longer run (15). The discriminator there is the motion around the run: flow standard deviation 8.578 for the freeze against 0.000 for the static control.

Warping is invisible under a pan#

Comparing each pixel's flow to the frame's median was built first and does not work. Under the pan a 6-pixel warp raised global incoherence by 2.6% and lowered the outlier fraction, because a subject and a background at different depths already disagree with the frame median.

Comparing to a local neighbourhood mean is the right measure, and it is decisive when the camera is still:

Case on a static camera mean local incoherence ratio to baseline
no warp 0.00041 1×
warp, 2 px amplitude 0.01078 26×
warp, 4 px amplitude 0.02252 55×
warp, 6 px amplitude 0.03325 81×

Under a pan the same warp moves it from 1.170 to 1.215 — 4%, indistinguishable. So warpingMeasurable is reported per frame pair, mean flow above 1.0 disqualifies a warp claim, and detectWarping refuses on a moving shot rather than returning a number nobody can interpret. That refusal becomes an entry in the report's unmeasured list, because a report with six clean detections and a silently skipped seventh reads as seven clean detections.

The rest#

  • Motion smoothness is the second difference of flow magnitude, so a fast steady move is smooth: the pan's maximum acceleration is 1.025 against 18.48 for both the freeze and the judder.
  • Background instability needs the two regions measured apart — 11.4× against the pan's 0.88 — or a moving camera looks like an unstable background.
  • Flow outliers are pixels moving more than a whole pixel differently from their neighbours. An absolute threshold, so it means the same thing in every frame, unlike a multiple of the frame's own median (which flagged 42–45% of every sequence, including the static one's zero-motion frames, when it was tried).

Reproducing#

text
python tools/human-video/evaluator/make_motion_sequences.py --out /tmp/motion
python tools/human-video/evaluator/motion_quality.py --sequence /tmp/motion/judder \
  --subject-box 200,40,220,190

Per-item status this evidence supports#

Item Was Is
182.C.19.06 not-started done; seven of eight separate cleanly, warping only on a near-still camera and gated as such