How to Tell If a Video Is Too Shaky — and Why a Pan Isn't Shake

Ask people to rate a pile of ordinary phone videos and one thing pulls the scores down more than almost anything a quality report usually measures: the picture will not keep still. Bitrate, sharpness and exposure can all be fine and the clip still reads as amateur.

We measured it. On KoNViD-1k — 600 real Flickr videos, each rated by viewers — the usual technical checks together barely tracked what people thought. A single shake measurement, on its own, did.

Movement is not the problem

The obvious way to measure shake is to measure how much the picture moves. That fails immediately: a slow pan across a landscape moves the whole frame every single frame, and nobody calls it shaky.

What a viewer reads as shake is a change in movement. A pan moves at a steady speed — left by the same amount each frame. A hand-held camera moves left, then a little right, then up, then left again. So the measurement is the frame-to-frame acceleration of the picture, not its speed:

1. Estimate how far the whole picture shifted between two consecutive frames.
2. Do it again for the next pair.
3. The difference between those two shifts is the jolt.

A perfect pan scores close to zero. A tripod shot scores zero. A walk with the phone held out scores high.

Three things that must not be called shake

Fast action. A fight scene or a whip pan is not one global movement — half the frame goes one way and half another. Any estimate of “how far the picture moved” is a guess there, so low-confidence measurements are thrown away rather than counted. Before that rule, a professionally shot film fight scene was called shaky off those guesses.

A cut. Two frames either side of an edit are different pictures; the “movement” between them is meaningless. The same confidence rule drops it.

A frame-rate conversion. Convert 30fps footage to 25 and one frame in six goes missing. Every time it does, the picture jumps twice as far as usual — along the same path. That is judder, not shake, and the advice is the opposite: a tripod will not help, a proper frame-rate conversion (or shooting at the delivery rate) will. When the big jumps are mostly speed changes along one direction, the report names it Uneven motion instead of Camera shake.

What the numbers mean

The measure is reported as a percentage of the frame width — how far the picture typically jolts between frames:

These lines come from the KoNViD distribution itself: the median real video sits at about 0.10%, and the shakiest tenth above 0.38%.

Fixing it

• At the shoot: a tripod, a gimbal, or the camera’s own stabilisation setting. Leaning against a wall halves most hand shake for free.
• After the shoot: stabilisation in an editor works by cropping in and shifting each frame to cancel the movement. It costs resolution and can add a wobbly “jelly” look on fast shake, so apply the minimum that makes it watchable.
• If the report says Uneven motion: re-export at the frame rate the footage was shot at, or use a converter that blends or interpolates frames instead of dropping them.

You can run any clip through the video quality checker — it reports camera shake alongside bitrate, sharpness and exposure, and marks the seconds where it is worst.